The extracellular matrix (ECM) regulates biological pathways crucial for homeostasis [1]. n solid tumours, ECM–cancer cell interactions alter the microenvironment, influencing tumour progression, metastasis, and treatment response [2] [3] [4] [5]. Significant the Tumoural-Micro-Environment (TME) alterations include excessive accumulation of type I collagen and increased crosslinking, leading to higher stiffness [1] [3]. Magnetic Resonance Elastography (MRE) is widely used to assess tissue biomechanics—commonly in liver fibrosis staging [6]— and is gaining interest for predicting chemotherapy response [7]. Biomechanical modifications of the ECM in the context of cancer are currently receiving significant attention both in vitro and in vivo [5] [8] [9]. So far, MRE on tumour organoids/spheroids has never been investigated [9]. In the context of tailored medicine, developing a 3D culture model, from patient’s biopsy samples, that is predictive to their response to treatment could be relevant. Here we use Magnetic Resonance Elastography (MRE) to locally characterize the biomechanics of a cancer cell spheroid within a collagen matrix.

Matrix preparation Rat tail collagen type I was mix with ribose (200mM final concentration) and allowed crosslinking for 30 minutes at 4°C. Glycated collagen solution was mixed with OptiMEM, HEPES, fibronectin and neutralized with 1M NaOH (collagen final concentration: 3 mg/mL). Spheroids generation and embedding HepG2 cells were grown in 2D until 60% confluency. After trypsinization, 5000 cells/well were seeded in a U-bottom 96-well plate and cultured for 2 days. Then, spheroids were embedded in the collagen matrix: round shaped spheroids were selected with a microscope and mixed to the collagen matrix before plating. The collagen gelled at 37°C for 30 minutes before adding culture media. Additionally, 3% agar gels were prepared from which agar phantom organoids samples were cut into 500 microns pieces. These agar organoids were equally embedded in the collagen matrix. 200uL droplets of hydrogels were plated in a 35mm  petri dish, resulting in a dome geometry of 10mm x 1mm. MRE acquisition and post processing Data were acquired on a 7T MRI preclinical scanner, using a spin echo MRE sequence with 35 microns resolution in the readout direction (upward direction represented by the red arrow in Figure1), 0.3mm in phase-encoding direction, and 0.5mm slice thickness. Shear waves were induced contactlessly from the bottom of the petri dish, generating planar waves in one spatial direction (Figure1B&Figure2B). HepG2 and agar spheroids were localized within the magnitude image (Figure 1C, D, Figure 2 C, D). Figure 3 shows how the biomechanical properties of the spheroids with their microenvironment were quantified: from the representation of the sine function (Figure 3B) of the wave propagation along the red line of the Figure 3A, sine/cosine function were locally fitted to their imaginary and real part of the wave field using an exponential suppression. The fitting was segmented in three parts: 0-0.6-1.2-1.8mm (Figure 3). Thereby, stiffness could be quantified within a perimeter of 0.9mm laterally to the spheroid and until ~1.8mm longitudinally covering the entire matrix (Figure1D&Figure2D).

Figure 4A shows a wave velocity within the tumour spheroid of 0.34 m/s, slowly decreasing to 0.24 m/s within a vicinity of 600 microns (scale C1 showing the wave velocity in m/s) until reaching its nominal value, with increasing the distance to the HepG2 spheroid (scale C2, showing the relative shear wave velocity normalized to HepG2 spheroid). This microenvironment modulation is not seen in the agar spheroid (Figure 4B), where the shear speed drops immediately to its nominal value.

Shear speed maps (Figure 4) indicate a high modulation of the tumoural spheroid microenvironment after 2 weeks of culture within a radius of 600 microns. This modulation is shaped as a stiffness gradient around the HepG2 spheroid. In contrast, shear speed dramatically drops around the agar spheroid. Stiffening of the environment is a well-established hallmark in cancer progression, primarily driven by the crosstalk between tumour cells and the TME, which involves extensive remodelling of the ECM and overexpression of collagen [3] [4] [5] [10] [11]. These results go hand in hand with recent results where tissue mechanics is capable to identify very early during neoadjuvant chemotherapy in breast cancer resistance due to changes in biomechanics [7].

Our work has investigated the feasibility of MRE on in vitro models. Here we present a concrete application of micro-MRE on a 2 weeks tumoural spheroid. These encouraging results and the lack of non-invasive methods [9] [12] [13] to investigate biomechanics in vitro motivate us to improve micro-MRE and apply it to the field of tailored medicine with transposable biomarkers.

Q-space trajectory imaging (QTI) [1] estimates the diffusion tensor distribution from multiple compartments [2] using spherical (STE), linear (LTE) and planar (PTE) b-tensor shapes to assess anisotropy [3-4]. The encoding spectrum should contain similar frequencies across axes to ensure restricted diffusion measurement at the same diffusion time for each axis and across encodings as their combination should result in a diffusion tensor distribution, specific for one diffusion time [5-7]. To achieve the desired b-tensor shape and encoding spectrum while simultaneously accounting for gradient system limits prolongs the diffusion encoding gradients. This is especially challenging for measurements of white matter tissue at ultra-high field due to the faster T2 signal decay. In this work, the diffusion gradients of three established QTI methods were optimized for microscopic anisotropy measurements at 7T accounting for shape and encoding spectrum of the b-tensor. These methods were combined with a single shot spiral readout to minimize TE. In vivo QTI-metric maps of the three methods were compared with regard to encoding spectra differences, which were expected to result in metric maps at different diffusion times.

One healthy subject was measured on a MAGNETOM 7T Plus scanner (Siemens Healthineers AG, Germany) with a 32/1 ch receive/transmit coil (Nova Medical), and a maximum gradient strength of 70mT/m. Three diffusion waveforms were inserted into a single shot spiral spin-echo sequence [8] implemented with Pulseq [9] (Figure 1A). Diffusion waveforms were designed according to the qMAS [10-11], gDOR [6-7] and FAMEDcos [11-12] approaches, which probe different encoding spectra [6-7], and were optimized in the following ways: qMASmod [11] (Figure 2A) was modified with additional correction factors considering a finite gradient slew rate ensuring STE. Using MATLAB’s “fmincon” [13-14], STE waveform parameters were optimized (Figure 1B) to minimize TE with regard to scanner limits and two (one) STE gradient axes were used for PTE (LTE). gDORrefoc (Figure 2B) was adapted from Ref. [6-7]. The one-dimensional gradient g1D(t) [6-7] was optimized using MATLAB’s “generic algorithm” [13,15] leading to a spin-echo with minimal TE (Figure 1C). The double oriented frequency ratio n and the total angle of rotation ΔΨ2 [6-7] were optimized to minimize frequencies near 0Hz. PTE and LTE were optimized separately and STE was excluded to reduce TE. For FAMEDcos [11] (Figure 2C) sinusoidal gradients were exchanged with trapezoidal gradients for STE [12] reducing TE and two (one) STE gradient axes were used for PTE (LTE). Image reconstruction [16] was done with the PowerGrid toolbox [17] using an expanded encoding model [18], combining B0 and coil sensitivity maps with field camera data (Skope, Zürich). QTI-metric maps were computed on preprocessed data with the QTI± algorithm [19-21]. The metric maps were compared to each other within the splenium (parallel fibers) and the Anterior Corona radiata (ACR) (crossing fibers) to investigate encoding spectra differences of the three methods. Therefore, metrics as well as FSL atlases [22-24] of the regions, were registered to an MP-RAGE image. Violin plot statistics, voxel-wise differences and Cohen’s d [25] were computed.

qMASmod resulted in higher µFA and lower Cmd across the white matter compared to gDORrefoc and FAMEDcos (Figure 3). The comparison between qMASmod and the other approaches in Figure 4 showed a high effect of Cohen’s d (d>0.8) with large voxel differences for µFA and Cmd in the ACR and a slightly reduced effect in the splenium. In contrast, gDORrefoc and FAMEDcos led to lower d in these cases. The MD and FA in the splenium led to d<0.2 (low effect) for all cases with differences centered around zero, whereas qMASmod led to lower MD values and higher FA values in the ACR (d≈0.5).

The in vivo measurements showed that QTI-metric maps can be obtained at 7T by combining optimized diffusion gradients with a spiral readout minimizing TE. A maximum b-value of 1400 s/mm² was chosen instead of 2000 s/mm² [3] to not increase TE even further, but still resulting in reasonable anisotropies [1]. The difference of µFA and Cmd in the splenium and ACR could be explained by the difference of the encoding spectra. While qMASmod measures QTI-metrics at very long diffusion times leading to higher µFA and lower Cmd, FAMEDcos and gDORrefoc measure the metrics at shorter diffusion times. These results agree with the literature [6-7]. The difference is more visible for the region with many crossing fibers (ACR) compared to the region with many parallel fibers (splenium) and not that strong for the macroscopic metrics (MD, FA), at least in the investigated regions.

QTI-waveforms were optimized and inserted into a spiral spin-echo sequence to obtain feasible acquisition times at 7T. The variation in QTI-metrics across sequences suggests that encoding spectra affect metrics due to different diffusion time.

Myocardial T1 and T2 mapping sequences have been widely used in the past decade to better assess tissue scar and oedema. A novel technique called T1rho mapping had been developed recently (1) to study low frequency phenomenon such proton exchange between water and macromolecules. It is suggested that T1rho would be able to provide contrast between myocardium and interstitial fibrosis and other heart diseases involving large molecules. The key benefit would be in the end to provide myocardial lesion detection and quantificatione without the need of contrast agent injection.

Two patients underwent standard MR clinical procedure on a 1.5T Sola MRI (Siemens Healthineers, Erlangen) consisting of conventional cines, T1, T2 mapping, dynamic perfusion and delayed enhancement protocols. A stack of T1rho slices acquired under breath-hold in short axis view was added to the protocol before contrast agent injection. The T1rho sequence parameters were: FOV 360x290, voxel size 2.5x1.9x8mm, T1rho preparation module used 4 Spin-Lock pulses and 2 refocusing pulses. Myocardial T1rho maps were generated using elastic coregistration and a 2-parameter curve fitting process using five different Spin-Lock times: 0, 10, 20, 35, 50ms.

On the first patient with suspected acute myocarditis, as shown on fig.1 there is a very good agreement between delayed enhancement images and T1rho maps, with T1rho values elevation in the area of LGE (78 msec vs. 45 msec). On the second patient with suspected cardiomyopathy (Fig. 2), there is a more discrete enhancement on both delayed enhancement and T1 rho map on the inferolateral wall. T1rho values increase from 44 msec in normal tissue vs. 55msec in abnormal myocardium.

T1rho maps exhibit good correlation with delayed enhancement images. This new contrast seems to be very sensitive to either water content or tissue scar. As questions arise whether gadolinium-based contrast agent can be harmful to human health and environment, this technique provides a useful tool for a number of patients (children, pregnant women, patients with renal failure). The relative low resolution compared to a standard delayed enhancement protocol in the other hand could be a drawback for systematic use.

Cortical development shapes cognitive function and brain health, with abnormalities linked to neurological disorders [1,2]. In early childhood (1.5–6 years), associative neuron maturation drives key cognitive shifts [3]. Studying the late fetal or preterm period, marked by dendritic differentiation and neuronal aggregation, remains challenging in vivo. Brain organoids derived from induced pluripotent stem cells (iPSCs) offer a promising in vitro model, replicating key developmental features while bypassing ethical concerns [4]. Here, we used diffusion MRI (dMRI) to assess microstructural maturation in cortical organoids generated from healthy donor hiPSCs (NeuroNA Human Cellular Neuroscience Platform). We applied diffusion tensor and kurtosis imaging metrics (DTI/DKI) [5] to evaluate maturation [6] and examined diffusion time sensitivity, highlighting organoid-based dMRI as a non-invasive method for studying early cortical development.

Nine brain organoids (~2 mm) were cultured for 4.5 months and combined into three assembloids (~4 mm). Organoids were fixed in 4% paraformaldehyde for up to 2 hours (based on size and ~1 mm/h diffusion rate), with swirling for uniform fixation, then washed in PBS and stored. Assembloids were scanned at CIBM (Lausanne, Switzerland) using a 9.4T Bruker MRI CryoProbe, stabilized in agar within syringe holders. Diffusion MRI was acquired with a PGSE EPI sequence using b-values [1, 2, 3.5, 5, 7] ms/µm², directions [12, 16, 24, 30, 40], and Δ = [15, 26, 38] ms with δ = 4.5 ms; four b=0 images per Δ. Other parameters: TE/TR = 54 ms/2.4 s, 0.2-mm resolution, 0.4-mm slice spacing; 50 min scan per Δ. Data were denoised with Patch2Self, corrected for Gibbs ringing (DIPY), and adjusted for susceptibility, eddy currents, and motion (FSL topup and eddy)[8]. DTI (b=1 ms/µm²) and DKI fits were done per Δ using DIPY. MD time-dependence was analyzed with the CEXI model [9], simulating exchange between intra- (spheres) and extracellular compartments using synthetic b=1 ms/µm² data.

Nine cerebral organoids were successfully cultured to 4.5 months. Figure 1A shows a set of three organoids assembled for imaging. The dMRI protocol yielded high-quality signals across all samples, with b=0 and b=7 ms/µm² images shown in Figures 1B and 1C. Based on visual contrast, inner and outer regions were manually segmented. The DKI model was applied separately to datasets acquired at different diffusion times (Δ). Figure 2 shows representative FA, MD, and MK maps for Δ = 15 ms and Δ = 38 ms. FA was generally low and decreased with increasing Δ, with slightly higher values in the outer ring. MD was higher in the inner region and increased with Δ in both regions. MK was elevated in the inner ring and decreased with diffusion time. Voxelwise distributions (Figure 3A) confirmed these regional trends. Figure 3B illustrates a clear increase in mean MD with Δ, consistent across DTI and DKI estimates. To explore the unexpected rise in MD—opposite to the typical decrease seen in white matter [10]—we performed simulations using the CEXI model. With impermeable membranes, MD remained constant or decreased with Δ (Figure 4A). Introducing membrane permeability led to an increase in MD, with the extent depending on cell radius and permeability (Figures 4B and 4C). Our simulations suggest that the increased MD in real data is a hallmark of transmembrane water exchange.

Across all samples, FA remained low (<0.15), indicating poor microstructural organization. The outer ring showed slightly higher FA than the inner zone, suggesting early development of elongated pyramidal and callosal neurons, as well as astroglial processes among rounded progenitors. At longer diffusion times, FA declined further due to increased permeability, allowing water to move across aligned and non-aligned structures, reducing anisotropy [11]. MD was consistently higher in the outer ring, reflecting the accumulation of oligodendrocyte precursors, immature interneurons, and progenitors in the cortical-plate-like region during months 4–5 [11].MD increased with diffusion time (Δ), consistent with CEXI model predictions of water exchange between cell bodies and the extracellular space. Conversely, MK decreased with Δ, as higher permeability and longer diffusion times blur compartmental distinctions, leading to more homogeneous diffusion and reduced kurtosis.

Time-dependent diffusion MRI sensitively captures microstructure in human cortical organoids. The diffusion time-dependent increase in mean diffusivity, supported by CEXI model simulations, indicates active water exchange between compartments. This exchange-driven MD pattern offers a non-invasive marker of early cellular development, highlighting the value of dMRI and organoid models in studying human cortical formation.

Quantifying the myelin sheath radius of myelinated axons in vivo is important for understanding, diagnosing, and monitoring various neurological disorders. Despite advancements in diffusion MRI (dMRI) microstructure techniques, models specifically designed to estimate myelin sheath radii remain unavailable. Recent studies suggest that it is possible to acquire dMRI data significantly weighted by myelin water using T1-based double inversion recovery (DIR) and T2-based relaxation selective measurements with short echo times (TE). Thus, it is the right moment to develop new microstructure dMRI models for this habitually neglected white matter (WM) compartment.

In this proof-of-concept theoretical study, we propose two novel dMRI models that characterize the signal from water diffusion confined to cylindrical surfaces, approximating myelin water diffusion (1). In the first, more general model, we derived an exact analytical expression for the dMRI signal using the diffusion propagator formalism based on the narrow pulse approximation. The second model employs a Gaussian approximation with time-dependent radial diffusivity, offering a simpler analytical relationship. We also developed approximate signal expressions for PGSE protocols with trapezoidal and rectangular diffusion gradients, extending beyond the narrow pulse assumption. We derive the spherical mean signals for the proposed models, which conveniently eliminate fiber orientation and dispersion effects. For the model based on the Gaussian approximation, the radial diffusivity depends on the cylinder radius r as, D⊥ = (r2/2t) [1-exp(-Dt/r2)], where t=Δ+δ is the total diffusion encoding time of the PGSE sequence and D is the axial diffusivity parallel to the main cylinder axis. These models are further extended to account for multiple concentric cylinders, mimicking the layered structure of myelin. Additionally, we introduce a method to convert histological distributions of axonal inner radii from the literature into myelin sheath radius distributions and derive analytical expressions to estimate the effective myelin sheath radius expected from these distributions (e.g., using a single-cylinder model). Figure 1 shows a schematic transverse section of a myelinated axon. We approximate the diffusion process along this spiral trajectory as diffusion within a series of impermeable concentric solid cylinders separated by infinitesimal water-filled gaps.

Monte Carlo (MC) simulations conducted in cylindrical and spiral geometries were carried out to validate the models employing a maximum gradient strength of G=500 mT/m. Figure 2 shows an example of the agreement among the MC-generated signals and those predicted by the general analytical model and the Gaussian approximation for different b-values and a range of cylinder radii. Additionally, we found a significant agreement between the effective myelin sheath radii estimated from reported histological distributions and the effective radius obtained by fitting a single-cylinder dMRI model to the MC signals generated from these distributions; see Figure 3 (for more details see (1)).

Despite our promising results, some limitations of this study must be addressed in future work. While the analytical models closely match MC simulations for various experimental conditions, discrepancies arise at high b-values and large diffusivities. These inaccuracies are due to the assumptions behind the narrow pulse approximation, which, despite the correction framework we introduced for more general PGSE sequences, remains an approximation valid primarily for Gaussian diffusion (low b-values). All results presented in this study are based on synthetic signals derived from the proposed analytical models or MC numerical simulations. Validation with real dMRI data, including histological analyses of various brain regions, is crucial for future work.

The proposed models hold the potential for estimating the effective myelin sheath radius from real dMRI data. For example, in diffusion-T1 experiments using inversion recovery sequences effectively isolating signals arising from myelin water, as outlined by (2), our models could be applied directly to fit the measured data. Similarly, for acquisition sequences where signals from other compartments are not entirely suppressed—such as in the diffusion-T2 hybrid sequences proposed by (3,4) or the magnetization-prepared dMRI sequence by (5)—our models could be integrated into a multi-compartment dMRI model to selectively fit the myelin water component. These approaches could be applied to both ex vivo and in vivo data, employing scanners equipped with strong diffusion gradients, where recent advances (6–8) make it feasible to enhance myelin water dMRI signal contribution by reducing the TE.

Liver T1 mapping is a helpful MRI technique to depict diseases like fibrosis or cirrhosis [1]. Whole-liver T1 mapping has been performed with 3D Variable Flip Angle sequences but may exhibit higher bias and lower repeatability/reproducibility compared to 2D approaches, like Look-Locker [1,2]. 

However, current 2D T1 mapping methods are limited to the acquisition of only 1 to 3 slices in one breath-hold.

One way to circumvent this problem is the use of the Simultaneous Multislice (SMS) method where several slices are acquired at the same time and unfolded during the reconstruction using the sensitivity coils information. This method has successfully been used for cardiac T1 mapping [3]

 

The MP2RAGE sequence is another method known for providing reliable T1 maps. A 2D multislice version of the MP2RAGE has previously been proposed for rapid T1 mapping in mice [4]. Six slices could be acquired in only 9 seconds by cleverly using “dead times” of the sequence to interleave the inversion and readout of the other slices.

This abstract describes how we brings all these ideas together using open-source tools for sequence development (Pulseq [5] and KomaMRI.jl [6] ) in order to develop a SMS 2D-MP2RAGE and achieve a rapid full-coverage liver T1 mapping in a single breath-hold on a 3T siemens.

A Multi-Slice MP2RAGE sequence previously developed on a Bruker 7T scanner [4]  was transferred to a 3T Siemens human scanner and modified to include SMS pulses and calibration lines (Fig 1).

For sequence development, simulation and validation of the developed sequence, we used Pulseq and KomaMRI tools. SMS excitation (sinc shape) and inversion (secant hyperbolic shape) RF pulses were generated by the sum of 2 RF shapes with an additional temporal phase term. Additional gradient echo calibration lines (24 for each slice) were added at the end of the MP2RAGE acquisition.

The main parameters of the sequence were : TI1/TI2 = 1000/3000ms, flip angle TI1/TI2 = 7°/7°, TR = 5ms, Echo Train Length = 128, SMS Slices = 6 (Total unfolded slices = 12), Slice thickness = 8 mm, Slice gap = 4 mm, matrix = 128 x 128, FOV = 320x320x144mm,  in-plane spatial resolution = 2.5 mm x 2.5 mm, total acquisition time = 8.66 (MP2RAGE) + 1.44 (calibration) = 10.1 seconds.

 

A direct SMS reconstruction was applied using the coil sensitivity maps computed from the calibration lines using the espirit algorithm from MRIReco.jl [7] then the quantitative maps were generated using a lookup table calculated without steady-states.

Pulseq sequence format can be used for both the acquisition as well as for simulation with KomaMRI.jl. This feature is especially useful to validate that the sequence and reconstruction algorithm are working before planning any experiments. 

Simulations with KomaMRI were used to: First, validate the slice profile selectivity of SMS excitation and inversion pulses. The second simulation on a numerical brain phantom validates that we were able to generate the FOV/2 offset of one of the slices.

Of note, simulation helps to debug a lot of aspects of a sequence but requires well defined simulation parameters. For example, in Figure 2 first row, the” slice 2” was not the same one as on the SMS images. After investigation, it was an issue with the simulation rf timestep parameter which was too high to correctly capture the linear phase necessary to generate the slice offset, and results in aliasing of the slice excitation. Decreasing the timestep from 50us to 25us solved the issue (Fig 2 second row).

 

After the success of all the simulations, the acquisition was performed on a healthy subject during a breath-hold. The T1 maps, covering the whole liver,  were reconstructed after unfolding the slices and (Fig 4). The mean liver T1 values (781ms  +/- 50 ms) was consistent with literature.

Before applying the newly-developed sequence on humans, Figu 2 third row shows the unfolding of the SMS acquisition and validates that the reconstruction algorithm is working as expected.

Using an Open-Source framework for sequence development is particularly useful when the access to a scanner is limited like in a clinical environment. The implementation of a sequence sufficiently well described in a publication can be done in a few days. Simulation is crucial to validate each key step,  like the SMS RF pulses in our case.

The next steps will consist in validating the SMS-MP2RAGE sequence against gold-standard methods and investigate sensitivity to motion, B1 and B0  field inhomogeneities. 

In order to further accelerate the method without changing the echo train length, parallel imaging will be implemented and the number of calibration lines will be optimized to keep the total scan acquisition as low as possible.

In this work, we demonstrated the feasibility of developing a rapid, full-coverage liver T1 mapping sequence using open-source tools. These tools facilitated in this work   the enhancement of a standard MP2RAGE sequence with advanced acceleration methods like SMS.

Magnetic Resonance Elastography (MRE) is a non-invasive imaging technique that measures the biomechanical properties of in vivo brain tissue [1]. Brain stiffness, a potential diagnostic biomarker, decreases with age and prematurely in neurodegenerative diseases [2]. However, brain MRE typically uses a narrow mechanical frequency range from 30 to 50 Hz, limiting viscoelastic dispersion analysis. A wideband MRE approach with a broader frequency spectrum aids multi-frequency dispersion analysis and potentially enhances our understanding of age- and disease-related microstructural tissue changes associated with biomechanical property changes.

This observational study recruited 15 young adults (mean age: 29.7 ± 4.1 years) and 10 older adults (mean age: 56.5 ± 3.8 years). A new wideband MRE setup was developed, incorporating two actuator designs: a cushion actuator positioned laterally beside the head for low frequencies (5-20 Hz) and a plate actuator [3] positioned beneath the head for higher frequencies (20-50 Hz). Both actuator systems can be simultaneously fitted into a standard MR head coil, enabling continuous shear wave induction across 5-50 Hz. MRE data were acquired using a spin-echo echo-planar imaging (SE-EPI) sequence with flow-compensated motion-encoding gradients (FOV=220x220x18 mm³, Matrix=112x140x15, TE=56 ms, TR=1400 ms). Shear wave speed (SWS) maps were processed using the kMDEV inversion algorithm [3]. A frequency-SWS dispersion analysis was performed by fitting a spring-pot rheological model to SWS data to assess age-related differences in brain tissue stiffness.

The wideband MRE setup successfully induced shear waves from 5 to 50 Hz. Frequency-resolved wave fields and shear wave speed (SWS) maps are shown in Figure 1. Frequency-averaged brain stiffness, measured by SWS, was lower in the older group (1.16 ± 0.03 m/s) than in the younger group (1.22 ± 0.04 m/s; p<0.001). SWS decreased linearly with age (SWS = 1.28 - 0.0021×age; R²=0.365, p=0.002), equivalent to a -0.20 % annual reduction, as shown in Figure 2. Spring-pot modeling on frequency-resolved stiffness maps showed that the stiffness modulus E decreased with age (R²=0.314, p=0.005), while the fractional exponent α increased (R²=0.236, p=0.019), as depicted in Figure 3.

The dual-actuator system effectively extended the frequency range of in vivo MRE to an unprecedented wide range of more than three octaves, enabling comprehensive multi-frequency dispersion analysis. Our wideband MRE approach provides deeper insights into age-related biomechanical changes of neural tissue compared to standard mono-frequency MRE. The observed age-related decrease in SWS and changes in spring-pot parameters suggest sensitivity to microstructural tissue degradation such as the loss of neural and vascular integrity, which is known to be associated with physiological aging. Although this study validated the technical implementation of wideband MRE, future research will explore its application in patient cohorts and interventional studies.

Wideband brain MRE, based on a dual-actuator design, robustly assesses brain mechanical properties in a wide frequency range from 5 to 50 Hz. Age-related differences in SWS and viscoelastic properties highlight the potential of the new MRE technique to detect microstructural changes, offering a valuable tool for studying brain aging and neurodegenerative diseases.

Blood-brain barrier (BBB) water exchange may serve as an early biomarker for Alzheimer's Disease (AD). Novel arterial spin labeling (ASL)-based BBB MRI sequences have the potential to become early non-invasive imaging biomarkers, especially because blood water as endogenous tracer may be sensitive to earlier and more subtle BBB changes compared to other contrast-based BBB imaging techniques. Novel multi-echo BBB-ASL sequences have already been shown to be reproducible and to be associated both with aging and AD in both animal models and humans. With the recent advent of larger BBB-ASL cohorts, it becomes possible to investigate to what extent BBB-ASL is associated with aging-related vascular pathology and AD-specific pathology. This study investigated the potential of a non-invasive multi-echo arterial spin labeling-based technique to assess BBB water exchange in relation to amyloid, cognition, and cerebrovascular burden stages.

Participants older than 50 years were selected from the Center for Lifespan Changes in Brain and Cognition (LCBC) and the Dementia Disease Initiation (DDI) cohorts. All participants were scanned on the same 3T Siemens Prisma scanner with a 32-channel head coil. Two multi-post-labeling delay (PLD) Hadamard-encoded (HAD) 3D GRASE pCASL sequences were used to estimate Tex and CBF: 1) single-echo time (TE) HAD-8 with a sub-bolus duration 400 ms, PLD [600:400:3400] ms, TE = 12.5 ms; 2) multi-TE HAD-4 with sub-bolus of 1000 ms, PLD [1500:1000:3500] ms, and 8 TEs [14.4:28.9:217.2] ms. Images were analyzed with ExploreASL 1.12.0beta; CBF, Tex and ATT were quantified with FSL-FABBER and assessed in the total gray matter (GM; partial volume>0.7). Amyloid status was defined as positive (A+) or negative (A-) from the CSF amyloid-beta 42/40 ratio (cut-off ≤ 0.077) or amyloid-PET by visual reads when available. Total WMH volume was quantified with Fazekas scores (0-3) by trained neuoradiologists. Associations with age, sex, amyloid positivity (A+ or A-), cognition, and cerebrovascular burden were examined. Regarding cognition, participants were categorized into three groups: cognitively normal (CN), subjective cognitive decline (SCD), and mild cognitive impairment (MCI) (Table 1).

Tex was lower in SCD and MCI compared to CN (Figure 1D), while CBF differed only between CN and MCI (Figure 1E). Tex also decreased with increasing Fazekas scores (1–2), while ATT increased in Fazekas 3 (Figure 1G-I). No association with amyloid positivity was found after age and sex adjustment (data not shown). When including all variables in the same model (Table 2, rightmost column) — age, sex, amyloid status, cognitive staging, and Fazekas scores—Tex remained significantly different between SCD and CN and between MCI and CN. In this model, CBF again showed significant differences between MCI and CN. Finally, ATT showed a significant increase in Fazekas 3 compared to Fazekas 0.

Our results indicate that CBF and Tex were associated with amyloid, cognition, and cerebrovascular burden stages, though amyloid correlations did not remain significant after adjusting for all covariates. Specifically, BBB water exchange, measured by Tex, is associated with cognitive staging and with Fazekas scores and seems to be changing with these biomarkers earlier than CBF and ATT. Whether BBB water exchange alterations were the cause or effect of cognition and cerebrovascular burden cannot be differentiated from these data.

In conclusion, our findings suggest that the observed BBB water exchange effects are not directly related to AD-specific pathology but seem to reflect more general aging-related neurodegenerative processes, including cerebrovascular damage. Investigating BBB water exchange across a broader spectrum of diseases and comparing multi-modal BBB imaging approaches to assess BBB dynamics, could elucidate its potential as a universal or disease-specific biomarker. Finally, these results underline the importance of assessing early cerebral microvascular pathology in aging and neurodegenerative diseases.

The placenta is vital for the long-term health of both fetus and mother[1]. Placental blood flow is associated with adverse outcomes such as preeclampsia and fetal growth restriction[2]. However, the capillary structure and blood flow characteristics in placental remain poorly understood[3]. While recent methods, such as separating maternal and fetal blood flows[4] and using T2* information[5] to extend the intravoxel incoherent motion (IVIM) model[6, 7], can capture additional information, these approaches cannot directly measure incoherent flow in capillaries. Other imaging techniques like Doppler ultrasound[8, 9] or MRI-based phase contrast angiography(PCA-MRI)[10] are either insensitive to microvascular flow or best suited for coherent flow patterns, thus, leaving a gap that this study aims to address. We introduce a novel approach to estimate capillary-scale blood velocities using simulated diffusion MRI(dMRI) data and machine learning(ML) directly from in-vivo dMRI data, essentially adapting microstructure fingerprinting[11-13] for IVIM-type models. We construct in-silico capillary structures, simulate flow, generate dMRI signals and finally train ML regressors to learn flow parameters. Our approach outperforms the baseline IVIM, offering deeper capillary flow insights.

We present a Monte Carlo simulator for dMRI signal generation from in-silico placental capillaries, then describe velocity estimation using ML and IVIM. Simulation Two geometries mimicked random isotropically-orientated placental capillaries(Fig. 1): (1) a 3D ring-ball structure of three circular channels, and (2) randomly oriented cylinders, enabling simulation of IVIM-like pseudo-diffusion[6, 7]. Monte Carlo simulations tracked spin motion under flow (following capillary centerline) and diffusion (randomly in 3D), with displacement[14, 15]:x(t)=√(6D∙dt)+v∙dt, where D is the diffusion coefficient, dt the time step, and v the local spin velocity. Outside capillaries, only diffusion was considered. Internal(S_i) and external(S_e) signals were computed separately. The final synthetic dMRI signal was: :S_sythetic=V_f∙S_i+(1-V_f)∙S_e, where V_f is the perfusion fraction[16]. Key ring-ball parameters: L=[10,80]µm (vessel length), d=[5,30]µm (diameter), D =[0.001,0.003]mm²/s, v=[0.001,1]mm/s, sampled via Latin Hypercube Sampling(1000 combinations). Random cylinders considered infinite length to prevent spin trapping with d, D, and v evenly distributed (10 samples within above ranges, yielding 1000 combinations). Noise (SNR = 20) replicated in-vivo conditions[17]. These combinations generated 1000 in-silica training dMRI signals per geometry. The test dataset had 60 velocity-varying combinations(L=100 µm, d=6, 10, 14µm, D=0.002mm²/s). Parameter Estimation We IVIM-fitted testing dataset. S_0 and D_initial came from monoexponential fitting of b>200 s/mm^2[18]. Velocities were derived from IVIM using the Einstein relation (D^*=Lv/6)[19]. Initial V_f, L and v used the simulation parameters(ground truth), unfairly advantaging the IVIM method. Two ML regressors, Random Forest(RF) (1000 estimators, max depth 40) and Multi-layer Perceptron (MLP)(layers: 64, 32, 16, 4; learning rate 0.005) in scikit-learn, were trained on synthetic data to predict L, d, D, and v. All parameters trained jointly. Trained regressors were applied to in-vivo scans (78 cases, 69 pregnancies, from publication[20, 21]) to predict v and D maps.

Fig. 2 shows capillary velocity predictions on the synthetic test dataset using RF, MLP, and IVIM, with/without noise. Fig. 3 shows predictions on in-vivo data from regressors trained on ring-ball structure only and both structures together. Fig. 4 displays velocity/diffusion coefficient trends during GA via MLP/RF.

From Fig. 2, ML regressors outperformed IVIM in predicting velocities. RF and MLP showed high accuracy in noiseless data(MLP: r≈0.99; RF: r≈0.98), while IVIM performed poorly(ring-ball: r≈0.84; random cylinder: r≈0.7). With noise, MLs remained robust(MLP: r≈0.94; RF: r≈0.93), whereas IVIM worsened (r<0.55). MLs showed limitations at extremes but surpassed IVIM, which oversimplifies flow into a single pseudo-diffusion term. Figs 3 showed training data diversity affected predictions, especially for MLP. Predicted velocities in Fig. 4 (0.3–0.9 mm/s) matched PCA-MRI findings(~0.6 mm/s)[10]. Both ML models showed a gestational decline in capillary velocity and diffusion coefficients, aligning with previous reports[22-24].

This study introduces a Monte Carlo and ML-based method for estimating placental capillary velocities, providing reliable predictions despite noise. Including more diverse structures enhances feature extraction, especially for MLP. Future work will refine the approach with more complex structures and validate using in-vivo PCA-MRI or ex-vivo perfusion imaging. Once validated, this method could non-invasively assess blood flow dynamics across organs in physiological and pathological conditions.

Gliomas are common and aggressive tumours of the central nervous system, with a five-year survival rate of only 5% (1,2). Aquaporin 4 (AQP4) plays an important role in glioma pathology and could serve as a prognostic biomarker (3–5). AQP4 is difficult to measure directly in vivo, but can be mapped indirectly with MRI via the water exchange rate (3). However, this demands diffusion MRI (dMRI) data acquired with non-conventional approaches such as free gradient waveforms (6). In this study, we evaluate whether dMRI reveals exchange-driven contrast in gliomas that complements standard imaging parameters such as apparent diffusion coefficient (ADC), kurtosis, and contrast enhancement.

Theory: We use the restriction-exchange framework (6,7) to analyse the MR signal, S, via ln⁡(S/S_0)=-b⋅E_D+1/6 b^2 E_D^2 K_T⋅(1-kΓ) #(1) where there are four microstructural parameters: S_0 (signal without diffusion weighting), E_D (mean diffusivity), K_T (kurtosis), k (water exchange rate) and two experimental parameters: b (b-value) and Γ (exchange-weighting) given by Γ=2/b^2 ∫_0^T t q^2 (t) q^2 (t+t' )dt' #(2) where q(t)=γ∫_0^t g(t') dt' is the q-vector and γ is the gyromagnetic ratio. Waveforms are designed with varying Γ but fixed sensitivity to restricted diffusion to enable selective sensitivity to exchange. Gradient waveform design: Two exchange-encoding gradient waveforms were optimised following (6) (Fig. 1A). Monte Carlo simulations were used to verify the exchange weighting properties of the waveforms (Fig. 1B). Participants: In this exploratory study, five participants with primary brain tumours (cases 1 through 5) aged between 56 and 74 were examined. Cases 1-4 were diagnosed with grade 4 glioblastoma and case 5 with astrocytoma. MRI scans: All participants were scanned on a 3 T MAGNETOM Prisma (Siemens Healthineers, Forchheim, Germany) using a research sequence (8). The diffusion imaging used b-values of 0, 1.3, 2.6, and 4 ms/µm2 in 10 directions with a spatial resolution of 2×2×5 mm3, TE = 138 ms, TR = 4.6 s, strong fat suppression and acquisition time of 5 min. One participant was scanned longitudinally at 4 timepoints. Data analysis: Acquired images were denoised (9), motion- and eddy-current corrected (10), and powder-averaged. Exchange-driven signal contrast was calculated using ΔS(b) = ((S_(low Γ) (b) - S_(high Γ) (b)))/(1/2 (S_(low Γ) (b)+S_(high Γ) (b)) ) #(3) where S_(low Γ) (b) and S_(high Γ) (b) are signals acquired with waveforms at Γ=10 and Γ=40 ms, respectively, at a fixed b-value. ResEx parameter estimates were obtained by fitting Eq. (1). Explorative ROI-based analyses were performed using BraTS (11) to compare exchange estimates in diseased and healthy tissues. Given the small cohort, results were interpreted qualitatively without formal statistical testing.

Figure 1 illustrates that the two waveforms are specific to exchange. Figure 2 shows that exchange contrasts manifest in gliomas at higher b-values, as expected since exchange is a higher-order time-dependent effect. The contrasts show different trends across cases. Gd enhancement is connected to increased exchange contrast. ADC is high in oedema and necrotic regions where cellularity is likely low. Kurtosis maps show opposite patterns. Exchange maps resemble the contrast patterns. Fig. 3 shows exchange estimates (k) in oedema compared with contralateral healthy white matter. Some cases show slight elevation in k in the oedema, others a marked increase, suggesting disrupted membrane integrity. Figure 4 shows T1-weighted images highlighting progressive enlargement of both the core and enhancing tumour over time. ADC is consistently high in the core, and kurtosis is high in the enhancing rim. Of note, exchange hyperintensity extends beyond the enhancing rim over time, indicating tumour infiltration into peritumoural tissue.

Elevated water exchange rates, as measured by dMRI with free gradient waveforms, were detected in all five glioma patients. Exchange maps revealed tumour heterogeneity not captured by ADC or Gd enhancement, indicating potential for non-invasive assessment of membrane properties. Elevated exchange in tumours may reflect increased membrane permeability, possibly due to upregulation of AQP4 (12) or disruption of the blood-brain barrier (13) although the latter is an unlikely explanation due to the high b-values we used. These findings have both methodological and clinical implications, since accounting for exchange improves kurtosis estimates, which have been shown to aid in glioma grading (14,15). While limited by a small sample size, this study provides a proof-of-concept for in vivo mapping of water exchange in gliomas.

In conclusion, dMRI allows non-invasive mapping of exchange in gliomas and adds information to standard imaging. Exchange is related to AQP4 expression and is thus a potential biomarker of tumour aggressiveness. Future studies with larger cohorts are warranted to evaluate the diagnostic utility of exchange imaging.

Turbo Spin Echo (TSE) imaging at 7T faces increased challenges due to shortened T2 relaxation times at 7T restrict the echo train length (ETL). At the same time, longer T1 relaxation times require longer repetition times prolonging time per shot. Spiral TSE sequences, which combine high signal acquisition efficiency with reduced number of shots and shorter ETLs, are therefore particularly attractive at 7T for achieving high-resolution imaging within shot scan times. However, higher field strengths bring the challenge of inhomogeneous transmit (B1+) fields. Especially TSE sequences suffer from it, which impair uniform excitation and refocusing and designing slice-selective pTx pulses is complex. Moreover, spiral readouts will be affected by B0 inhomogeneities, resulting in a more complex k-space filter as in Cartesian sampling. To address these challenges, we propose a 7T spiral TSE sequence based on the 3T spiral TSE sequence by Hennig et al. [1], incorporating shorter spiral segments to reduce B0 sensitivity and applying Direct Encoded Signal Control (DESC) [2] to mitigate B1+ inhomogeneities. DESC allows the use of standard slice selective RF pulses while optimizing the shim mode for each pulse individually.

We employed a spiral TSE sequence [1] (encoding scheme in Fig. 1) in fixed mode with a fat saturation pulse at the beginning with a fixed cp shim mode. The sequence was implemented with a short echo spacing of 6 ms to minimize off-resonance effects, an effective echo time of 12 ms and a total acquisition time of 0.42 s. The ETL is 68, including one dummy pulse. A 90° excitation is followed by a 40° dummy refocusing pulse and subsequent 40° refocusing pulses. Except for the fat saturation pulse (cp mode), all shim modes were optimized individually. For B1-shim optimization, we used the MR-zero [3] framework with a differentiable Phase Distribution Graph [4] simulation written in PyTorch, enabling gradient-based optimization [5,6]. The loss function is the RMS of the magnitude difference between the reconstructed and ideal target image, which is simulated with constant, mean relative B1+ value of the cp mode. The simulation phantom used a combination of BrainWeb [7] data parameterized with reasonable relaxation values and multi-channel B1+ maps and a B0 map measured at 7T using the DREAM approach [8,9]. A low-resolution (42×42) phantom and static shim initialization were used to accelerate optimization. We optimized the B1-shims for the first 17 pulses (1 excitation pulse, one dummy refocusing pulse, 15 refocusing pulses); subsequent pulses used static shim mode. The resulting shim modes were exported using the pTx extension of the Pulseq standard [10,11]. Image reconstruction was performed at a resolution of 256x256 using NUFFT with and without off-resonance correction to address for B0 influence during the spiral TSE acquisition. The measurements were performed on a 7T MAGNETOM Terra.X scanner (Siemens Healthcare, Erlangen) on one healthy subject after written informed consent and approved by the local ethics committee.

The B1-shim optimization took approximately 3 minutes for 100 iterations and results in different shim magnitudes and phases for each optimized pulse. The resulting pulse-specific B1+ fields are illustrated in Fig. 2 (transversal orientation) and Fig. 3 (sagittal orientation). Notably, the excitation pulse achieves a higher mean B1+ factor compared to both cp mode or static shimming. The final image exhibited improved homogeneity relative to static shimming method (Fig. 4). Receive fields are eliminated. Despite improvements, residual image artifacts remain visible in both cases.

It was shown the first spiral TSE MR image at 7T with a long ETL, effectively limiting the impact of B0 inhomogeneities, additional off-resonance correction at NUFFT reconstruction stage had minor influence on the image quality. By performing a DESC optimization, we optimized the shim parameters of individual RF pulses within the sequence leading to improved image homogeneity. The duration of the optimization process could be further reduced by selecting a smarter initial guess and/or utilizing idle scanner time for measuring other clinical sequences in parallel. Despite this improvement, residual image artifacts remain present. These could result from flow effects, or gradient imperfection and will be investigated in the future. The minimal effect of the off-resonance correction indicates the potential to use longer echo spacing, which will allow larger spiral segments per ETL and thus will lead to reduction of the overall echo train length.

We demonstrated the first spiral TSE MR image at 7T effectively limiting the impact of B0 and B1 inhomogeneities.

Myelin bilayer mapping is a recently developed method for quantitative mapping of macromolecular myelin content in the brain. Signals stemming directly from the myelin lipid-protein bilayer are captured at multiple TEs using dedicated short-T2 hardware[1-3] and the zero echo time (ZTE)[4] sequence variant with hybrid filling (HYFI)[5], and subsequently separated from other signal components via a fitting procedure[6]. A key feature of ZTE sequences is that the gradient is ramped up before RF excitation. To excite the full FOV, the RF pulse bandwidth (BW) must therefore match the image BW. Short-T2 imaging requires rapid spatial encoding[7], which is achieved through strong gradients, yielding high image BWs. Consequently, high-BW excitation pulses are needed, which have high specific absorption rates (SAR). This high SAR limits the achievable flip angle (FA) to below the Ernst angle, meaning that the SNR is SAR limited. Low SNR, in turn, is compensated by increased scan time. The multi-TE protocol is designed such that shorter TEs have higher image BWs. So far, the same frequency-swept[8] RF pulse, with BW matching that of the TEmin image, has been used across all TEs for consistent excitation. As such, every image has reached the SAR limit, SARmax. For frequency-swept pulses, FA ∝ 1/sqrt(pulse BW), i.e., lower pulse BW gives a higher FA, so lower RF power is needed to reach the same FA. Therefore, by instead matching the pulse BW to the respective image BW, SAR is reduced for longer-TE images. To benefit from this reduction, the SAR safety limits must be considered. These are defined as moving time averages[9]: SAR ≤ SARmax per 6 min window, and SAR ≤ 2⨯SARmax per 10 s window. Reducing SAR in terms of the safety limits thus boils down to decreasing the limiting average SAR, denoted SAR*, which can be realised by efficient distribution of the protocol SAR. Here, we apply this concept to myelin bilayer mapping by optimising the RF pulses for each TE, splitting the images into shorter scans, and interleaving the scans to minimise SAR*.

A protocol with eight TEs is used; scan parameters are listed in Table 1. The RF pulses were chosen through an optimisation procedure to match the FA of the reference pulse (i.e., that of the TEmin image) at minimal RF power. For the interleaving, each signal average is split into three scans of 70–85 s each (depending on the angular undersampling factor). Each scan needs 5 s of setup time (no SAR) and 3 s of dummy excitations (with SAR), leading to a total scan time of 53:31 without splitting and 57:40 with splitting. The scans are executed in an order that minimises SAR*, as determined by a custom algorithm that tries to reach the optimal SAR* reduction (i.e., that achieved if the total protocol SAR is spread evenly over the full scan time): It places one scan at a time (first trying those with highest SAR) and checks if the optimal reduction is met. If no scan can be placed, it lowers the expected reduction and tries again. Each set of three scans is reconstructed into an image, the images are registered to each other, and signal averages of the same image are combined. Otherwise, the processing is the same as previously[6], including k-space filtering and extension, bias correction, and model fitting to produce the quantitative myelin bilayer map. A healthy 57-year-old male volunteer was scanned over two sessions, once with interleaving of the protocol and once without.

The relative SAR of the protocol without and with interleaving is shown in Figures 1 and 2, respectively. It is clear that the 6 min window represents the limiting average. With interleaving, the SAR* is reduced by 44.3%, which, when adjusted for the increased scan time, corresponds to an effective reduction of 39.9%. Myelin maps for the data with and without interleaving are shown in Figure 3 for similar slices. The map quality is slightly worse for the data with interleaving, but this is expected to be solved with due fine-tuning on the acquisition side.

Assuming a linear relation between FA and SNR, the achieved reduction in SAR* could be used to increase the SNR by ~40%, which could, ideally, be traded for a ~44% reduction in scan time, i.e., a total scan time of ~32 mins. The choice of splitting factor is flexible, and can be set individually per image. It represents a trade-off between widening the optimisation space and increasing the scan time (due to larger overhead). The uniform factor of three used here was chosen for demonstration purposes and is not necessarily optimal. Similarly, the scan setup duration can essentially be set freely (above a lower limit). As such, there is still room for further improvements.

Myelin bilayer mapping was successfully performed at 39.9% lower SAR* with comparable map quality. In future, this SAR* reduction could be traded for increased SNR or reduced scan time. Similar SAR* reduction options may be available on clinical systems depending on the SAR-monitoring implementation.

Liver diseases are prevalent, and for their identification and characterization, imaging perfusion is essential [1]. Arterial Spin Labeling (ASL) [2] is a non-invasive technique used for measuring perfusion without the need for contrast agents, typically by magnetically labeling the blood flowing into the organ of interest. Velocity Selective ASL (VSASL) [3] is a specialized method that labels blood based on its velocity directly in the imaging region. This method is advantageous since it eliminates the need for additional time for blood to reach the imaging region. Moreover, due to the challenge of dual blood supply from the hepatic artery and portal vein, this method is especially beneficial for liver imaging [4]. In this preliminary study, three different VSASL preparation schemes are compared and the quantified perfusion values determined.

For the determination of the perfusion in the liver, an in-house developed velocity selective saturation sequence (Gaussian compRefoc) [5], a velocity selective inversion (VSI) [6] and a VSI sequence with sinc modulation of the nine RF pulses (VSI sincMod) [7] were compared. The images were acquired under time-controlled breathing to minimize motion-induced artifacts on a 3T scanner (vidaFit, Siemens Healthineers, Erlangen, Germany) using the vendor independent sequence development framework gammaSTAR [8]. For the quantification of the hepatic blood flow (HBF), the general kinetic model for VSASL [9] was used and a mean value over the whole liver determined. For a test-retest regarding the robustness of the sequence, the Gaussian compRefoc was tested in the same healthy volunteer (30y, female) in two additional exams. Following parameters were used: TR=6500ms, TE=23ms, velo = 2 cm/s in z-direction, TI=1200ms, PLD=100ms, velocity-selective insensitive control-module, meas=12 including a phase cycling of the refocusing RF pulses for all three versions. For the M0 acquisitions the same VSASL sequences with TI=10s and no background suppression pulses were used.

Figure 1 shows the calculated perfusion-weighted images (PWI) for the three different VSASL techniques in the liver. It can be observed that the Gaussian compRefoc sequence displays the most homogeneous perfusion signal throughout the liver. In comparison, perfusion signal can also be measured in the VSI and VSI sincMod sequence acquisitions, but some regions of the liver are low in signal intensity, therefore, the overall signal is not as homogeneous as in the Gaussian compRefoc sequence. The quantification of the hepatic blood flow (Table 1) shows that the quantified values are slightly lower than the global hepatic blood flow measured previously with pseudo-continuous ASL [10]. The test-retest results of the Gaussian compRefoc sequence (Figure 2) show that no homogeneous perfusion signal throughout the liver could be measured in the second measurement, and a signal progression from anterior to posterior, from positive to negative values, can be observed. However, the third measurement is comparable to the first one.

The acquisitions in Figure 1 clearly demonstrate the possibility of non-invasively measuring perfusion in the liver using all three VSASL techniques. Overall, the Gaussian compRefoc sequence displays the most homogeneous perfusion-weighted signal. This could be due to the fact that this sequence is specifically designed to be insensitive to B1-field inhomogeneities, which often occur in the liver [4]. The hepatic blood flow quantified with all three VSASL techniques yields reasonable values. The slight decrease could be explained by the different strengths of labelling the arterial or venous part due to varying blood flow velocities, which leads to a representation that isn't truly global perfusion. The test-retest results of the Gaussian compRefoc sequence demonstrate the sequence’s sensitivity to confounding effects. For instance, a suboptimal shimming process could have resulted in the decreased signal inside the liver in the second measurement. In the future, it will be necessary to acquire more data and enhance the robustness of the sequence, such as implementing an advanced and quick shimming process to homogenize the field, thereby stabilizing the signal.

In conclusion, the results demonstrate the ability to measure hepatic blood flow comparably using all three VSASL techniques. Among these, the specially-developed Gaussian compRefoc sequence yields the most reasonable results, assuming that the optimum of all system parameters is achieved during the preparation process. Future research aimed at stabilizing this preparation process is expected to enhance the robustness of the sequence.

Our group has recently developed a novel, fast multi-contrast GE-SE sequence [1] based on EPI keyhole techniques [2] to enhance spatiotemporal resolution and the number of echo samplings by incorporating two spin echoes for improved T2 quantification. The resulting sequence offers multi-parametric output in the form of simultaneous T2, T2*, vCBV and OEF estimation. The present work aims to extend the sequence variability to offer T1 quantification based on inversion recovery sampling.

The existing 10-echo GE-SE EPIK sequence [1] was extended by including initial slab selective inversion pulses of 200mm in width. Thereby, repeated IR-GE-SE acquisitions that sample different inversion times enable T1 recovery sampling [3]. It is important to note that, due to the specific EPI with keyhole methodology [4], all slices are acquired after each inversion pulse, therefore leading to different inversion time (TI) ranges with increments of (TR/slice number). The resulting echo times are 12, 24, 44, 56, 68, 80, 92, 112, 124 and 136 ms. Phantom and in vivo datasets of 10 slices with a TR of 2s were acquired with an acquisition time of 45s per TI for a spatial resolution of 1.9x1.9x3mm3. Eleven different TIs of 25, 100, 250, 500, 750, 1000, 1250, 1500, 2000, 2500, and 3000 ms were acquired. T2 and T2* maps were computed for each TI acquired in order to test the quantification stability.

The sequence diagram of the IR-GE-SE EPIK sequence is shown in Figure 1. T2 and T2* quantification based on the 10-echo data acquired at different TIs are in agreement with the literature and match the previous standard GE-SE EPIK quantification results. Exemplary maps are shown in Figure 2. The T1 maps show a good contrast quality, although the quantified values underestimated the expected values by 20-40%. The resulting T1, T2 and T2* distributions for the in vivo data sets are summarized in Figure 3a. Summarized mean values from the sphere phantom and WM/GM regions of the in vivo data are shown in Figure 4 alongside the expected values from reference acquisitions. Figure 3 (b, c) shows the resulting mean T1 value in a WM ROI for each slice to demonstrate the dependency of T1 on the actual TI range as the latter changes with the slice count. An overall increase in mean WM T1 can be seen across the slices with a plateau over the central six slices. In addition, Figure 5 shows the mean whole-brain T2 and T2* values obtained from each TI echo-series acquired. While a minor decrease in T2 is observed, T2* is robust and the quantification is stable for TI larger than 500ms overall.

While the map quality and T2/T2* robustness indicate the potential of extending the 10-echo GE-SE EPIK to T1 mapping, future work is needed on correcting the T1 underestimation. Potential reasons may be insufficient signal recovery or some cross-temporal effects arising from the EPIK sparse iterations to recombine single-weighted images. The T1 preparation method could be improved by using a TAPIR-like methodology [5] or slice-shuffling technique [6]. Moreover, a simple 2-point methodology will be tested in the future. All of these approaches should be directly compared to measured gold-standard values rather than expected literature values. In terms of T2 and T2*, the repeated acquisitions of different TIs may be useful for a suitable averaging method.

IR-GE-SE EPIK data offers quick access to simultanoues quantification of transverse and longitudinal relaxation times. Especially T2 and T2* show a great robustness and accuracy, while T1 contrast shows the potential towards an additional quantitative biomarker but requires future correction of its current understimation. The clinical relevance of a multi-parametric fast imaging sequence can be reached in future work by optimizing the IR-GE-SE EPIK perfomance.

Iron regulation in the brain, influenced by compounds such as ferritin, transferrin, and ferrous ions, is critical for normal brain function and is implicated in aging, neurodegenerative diseases. Recently, a quantitative MRI (qMRI) based approach was proposed to in vivo estimate the iron homeostasis [1]. In this method the dependency between two MRI relaxation R1 and R2* rate constants is estimated (r1-r2* relaxivity). In this work we validate this qMRI technology using an independent dataset of healthy young and old subjects and Parkinson’s disease (PD) patients, by comparing it to known histological results [2], [3].

qMRI scan from 87 subjects scanned in the Hebrew University [4]. From each subject R1, R2* and the r1-r2* relaxivity were estimated in 11 brain regions of interest (ROI). In vivo brain values (R1, R2*, r1-r2* relaxivity) were compared with ex vivo published quantification of iron compounds on the same ROIs.

Our analysis revealed consistent results, with in vivo qMRI scans correlating with ex vivo iron quantification. Specifically, the R2* highest correlation was to iron content (R-sq=0.76 p=8.1e-9) and the r1-r2* highest correlation was with iron mobilization (R-sq=0.68 p=1.6e-7). This reinforces the reliability of this qMRI-based method for assessing brain both iron fraction and iron homeostasis. A detailed analysis using ANOVA and Tukey's post-hoc tests to compare groups within the regions of ex vivo and in vivo values revealed inconsistencies in group differentiation between iron levels and R2* in the Globus Pallidus and Putamen, as well as between r1-r2* and iron mobilization in the Globus Pallidus and Caudate. These discrepancies may be attributed to the limited number of subjects.

This non-invasive qMRI method offers a valuable tool for monitoring brain iron homeostasis and could enhance the diagnosis and understanding of disorders related to impaired iron regulation.

Establishing the relationship between in vivo and ex vivo iron regulation opens the way for future research relating specific brain regions at disease states (e.g., PD) and the impact of specific genetic subgroups.

In musculoskeletal MRI, the T2 relaxation time is the most commonly estimated quantitative parameter, as it is an indicator of the extracellular matrix state. T2 mapping can serve as an important marker for therapeutic monitoring of tissue changes during disease progression or recovery [1,2] . The Dual Echo Steady State (DESS) sequence is of particular interest to perform such studies[3,4]. In addition to providing high contrast structural images, T2 maps can be obtained in short scan times. Nevertheless, two improvements are needed: 1) perform high spatial resolution in 3D without lengthening scan duration and 2) provide an accurate 3D T1 map to precisely estimate T2. To circumvent both issues, the DESS sequence was inserted into the 3D MP2RAGE sequence diagram [5], and the subsequent method, called MP2DESS, was evaluated on one volunteer at 3T.

The gradient echo encoding within the two echo trains of the MP2RAGE were replaced by a DESS encoding, enabling the reconstruction of two images per train: FID1, Echo1 and FID2, Echo2, respectively (Figure 1). The T1 and T2 maps were obtained by using two different and known T1- (MP2RAGE ratio proposed by Marques et al [5]) and T2- (DESS ratio proposed by Bruder et al [6]) sensitive image combinations to iteratively estimate each parameter, similarly to the method proposed by Heule et al [7]. A T2 map was estimated via look up table dictionary matching with the DESS ratio, which would then be used to measure a new T1 map, with a matching of the MP2RAGE ratio instead, and vice-versa. This two-way iterative estimation would then be run until convergence of the two parameters (Figure 2). Simulations demonstrated that this convergence was obtained rapidly, after 3 to 5 iterations. The signal dictionaries used for the look up tables were computed using the EPG formalism in the Sycomore toolkit [8]. Multiple sets of MP2DESS parameters were employed during scans as prior simulations had demonstrated that modifying certain parameters such as the flip angle, the inversion times and the total TR could have an important impact on the estimates. These parameters were modified with the constraint of keeping the acquisition time short (<=4 min). Acquisitions were carried out on a Siemens 3T PRISMA scanner on the knee of one healthy volunteer, with a 1mm isotropic resolution and the following parameters: Compressed Sensing acceleration factor = 4, ETL= 120 or 150, TE1/ TE2/TR=4/7/11ms, FA =10 or 20°, MP2DESS_TR=3200, 3800 or 4300ms, acquisition time between 3min33s and 4min01s. The T1 and T2 values were compared with their respective reference methods: the MP2RAGE (1mm isotropic, TE/TR=3/7ms, FA=7°/7°, MP2RAGE_TR=5s; ETL=125)[9], the DESS (1.2mm isotropic, TE1/TE2/TR= 5/10/15ms, FA=20°)[3,10] and a B1 map. The two reference sequences were also acquired with modified parameter sets to match the MP2DESS parameters. The relaxation times were measured in the cartilage, the meniscus and the muscle.

3D T1 and T2 maps were estimated from the reconstructed images of each MP2DESS scan. The meniscus, cartilage and muscle regions were clearly visible in both maps, in 3 out of 4 scans (MP2DESS 1,2 and 4) (Figure 3). Saturation of T1 estimates in the upper and lower regions of the images were due to low B1 values (<70%) not being included in the dictionary during the estimation. Mean T1 estimates for the three functional parameter sets were similar to the MP2RAGE reference values (<21% relative difference in the meniscus, <10% in the cartilage and <2% in the muscle, Figure 4). Mean T2 measurements in the cartilage and muscle returned <24% and <20% differences, respectively. T2 in the meniscus was estimated to be around 10ms in the MP2DESS scans, which reflects values found in literature [3], suggesting that the reference DESS scan was not optimal here (19.4ms). Modifying the parameters of the MP2RAGE and DESS sequences also impacted the value of both estimated quantities in a similar fashion.

The proposed sequence and T1 and T2 estimation method allows rapid musculoskeletal scans in 3D, with accurate measurements in most regions of interest depending on the choice of parameters. Interesting next steps could include reducing the voxel size for better segmentation of the cartilage, running Monte Carlo simulations to determine optimal parameters and performing test-rests scans measurements to assess repeatability of both quantitative estimates.

We have proposed a novel sequence that has the ability to measure reliable T1 and T2 estimates in the knee and benefits from rapid scan times and 3D isotropic voxel sizes. This method could be of great interest for reproducible musculoskeletal MRI, especially for the longitudinal monitoring of cartilage repair.

Resting-state functional MRI (rsfMRI) is a popular, non-invasive, indirect method to assess neuronal brain activity without specific stimuli[1,2] based on the Blood-Oxygen Level Dependent (BOLD) signal[3,4]. The BOLD signal has several features that offer unique and complementary insights into brain function and organization, which primarily focuses on functional connectivity (FC)[1]. Voxel-wise metrics such as amplitude of low-frequency fluctuations (ALFF)[5], fractional ALFF (fALFF)[6], regional homogeneity (ReHo)[7], entropy[8], and complexity[9] allow for reliable analyses of other features of the BOLD signal. However, some of these have had minimal usage in rsfMRI analysis compared to the more popular FC approaches[2,10], neglecting alternative characteristics of the BOLD signal. One problem with this, for example, is the lack of characterization of normative data to establish the localized rsfMRI baseline. Therefore, this pilot study investigated the normative range of various rsfMRI analyses in healthy subjects, establishing a foundational baseline for future research.

rsfMRI scans were retrieved from the open-access International Neuroimaging Data-sharing Initiative, available at https://www.nitrc.org/projects/fcon_1000/. rsfMRI brain scans from 50 participants (50:50 sex ratio, 20-40 years old) were used from the Nathan Kline Institute Rockland Sample. Data were collected on a 3T Siemens MAGNETOM Trio with the following parameters: resting state BOLD, Axial 2D, GE-EPI, 3mm isotropic, FOV=21.6cm, TE/TR=30/2500ms, 260 time points. Data were preprocessed as follows: (1) deletion of the first five functional volumes for magnetization equilibrium, (2) eddy current correction, (3) motion correction, (4) spatial smoothing with Gaussian kernel and FWHM=5mm, (5) 4D global normalization, (6) brain extraction and (7) registration to MNI space. Linear detrending and temporal filtering were applied according to the prerequisites of each analysis. After preprocessing, six rsfMRI analyses were performed: ALFF and fALFF, using the FFT-based approach[5,6,11]; ReHo, utilizing Kendall's coefficient of concordance (KCC)[7,12]; entropy, employing Approximate Entropy (ApEn) based on the average uncertainty of a time series[8,13,14]; fractal complexity through the Hurst exponent (H) using the rescaled range (R/S) method[15–17]; and chaos and regularity using the Lyapunov exponent[18–20]. Voxel-wise averages of each rsfMRI analysis were computed, along with their respective whole-brain histograms, to show the regional contrast and dynamic range of each metric. In addition, descriptive statistics were computed for each rsfMRI analysis, segregating by sex and region-of-interest (ROI), as previous research has reported sex and regional differences in some of these analyses[21].

A total of six rsfMRI analyses were investigated in a sample of 50 participants. Voxel-wise whole-brain averages of each of the six metrics are shown in Fig.1, displaying different contrasts between brain regions. Histograms depicting the value distribution of each rsfMRI analysis are shown in Fig.2, where it is seen that each analysis has a different distribution, mean, and dynamic range. Fig.3 shows the median and IQR (top), and the mean and SE (bottom), both are segregated by sex (male, female) and ROI.

In agreement with previous characterization, ALFF, fALFF, and ReHo voxel-wise averages showed maximal contrast between WM and GM, which were further supported by descriptive statistics[5,22]. In addition, this pilot study shows a similar behavior for H, ApEn, and Lyapunov exponent. Furthermore, the normative ranges and distributions showed distinct characteristics for each metric, with ALFF, fALFF, and ReHo having relatively higher standard deviations compared to H, ApEn, and Lyapunov exponent. Lastly, descriptive statistics suggest that ALFF, fALFF, ReHo, H, and Lyapunov exponent have higher GM values than WM, in contrast to ApEn.

This pilot study aimed to provide an initial characterization and baseline of six localized rsfMRI analyses in healthy participants. Although this has been performed for some analysis in independent research[21], to the authors’ knowledge, no study has attempted to characterize and compare these rsfMRI analyses. Nevertheless, in the future, it will be necessary to increase the sample size and expand to more datasets collected at other locations. Also, we will increase sample diversity and enable the use of more powerful statistical methods, testing for significant differences in these rsfMRI analyses across sex, age[16,23], and expand the number of brain regions included. Thorough characterization of these metrics could improve understanding of the healthy brain, which could help better understand brain diseases.

Phase-based magnetic resonance electrical properties tomography (MR-EPT) is an emerging technique to non-invasively measure tissue electrical conductivity (σ) from the MR transceive phase (φ0), via the Helmholtz equation [1]. Despite technical advancements, the repeatability of EPT remains largely unstudied. In this study, we investigated the repeatability of a noise-suppressing EPT reconstruction method we developed [2], optimised and applied to the human brain.

Data acquisition: Ten healthy volunteers (seven female, aged 23-30 years), recruited as part of a previous study [3], were scanned in two sessions one week apart, with three identical scans per session. Multi-echo 3D GRE images were acquired using a 3T Philips Achieva system. Sequence parameters are shown in Table 1A. EPT reconstruction: An in-house EPT pipeline [2] was applied to the denoised [4] complex GRE data. A total field map was obtained from a non-linear fit of the complex data over all TEs [5]. Residual phase wraps were removed using SEGUE [6]. The transceive phase, φ0, was estimated by predicting the complex signal at TE=0 and extrapolating the phase. Conductivity was reconstructed via the surface integral of the φ0 gradient, using 3D spherical kernels of radius 21 mm and 24 mm to perform differentiation and integration, respectively. An optimised magnitude- (Mag) and segmentation-based (Seg) weighting technique was employed, for noise suppression and edge preservation [2]. The magnitude image at the final TE was used both for Mag-weighting and to segment grey matter (GM), white matter (WM) and cerebrospinal fluid (CSF) with SynthSeg [7], for Seg-weighting. For Mag-weighting, the magnitude was adjusted by a free parameter, δ, which varied automatically based on phase noise levels [8]. Repeatability analyses: Repeatability was assessed both within-subject (including both intra- and inter-session) and between-subject, using some of the most popular and informative repeatability metrics (listed in Table 1B) [9-11]. For regional comparisons, the median conductivity was calculated within the GM, WM and CSF, excluding non-physical negative conductivity values which were considered erroneous. To enable voxelwise comparison, each subject’s conductivity maps were co-registered using FLIRT [12, 13].

The repeatability is illustrated in Fig. 1, which shows conductivity maps in a representative subject, across all six sessions (prior to registration). Regional variability is illustrated in Fig. 2, showing the median σ values in each tissue type, for all subjects across all six sessions. Greater variability was observed in the CSF compared to GM and WM. All repeatability metrics are shown in Fig. 3: average normalised root mean square error (NRMSE) across repetitions, and three kinds of standard deviation (SD; within-region, within-subject, and between-subject), calculated in each tissue type. All of these metrics were significantly higher in the CSF, compared to both GM and WM (p = 0.002).

The repeatability of our EPT method was reasonably high in the GM and WM regions, but was significantly poorer in the CSF. This is likely attributed to the higher rate of EPT errors in the CSF, generating negative and/or physically implausible conductivity values (> 10 S/m in the human brain). The reason for these errors remains unclear, but may relate to the increased anatomical complexity of the CSF regions, compared to more homogeneous WM areas [14]. Interestingly, in the grey and white matter, the between-subject SD was lower than the within-subject SD, suggesting that EPT is more repeatable across different subjects than within a single subject across repetitions. Future work will aim to minimise the CSF error rate, as well as compare repeatability using different EPT reconstruction methods.

In this preliminary study, we investigated the repeatability of EPT in the human brain, using a surface-integral MagSeg-weighted reconstruction method designed to suppress noise and preserve edges. Both visual inspection and quantitative analysis revealed that repeatability varies significantly between tissue types, with the poorest repeatability in the CSF. This work marks an important step toward EPT validation, and highlights minimising CSF-specific errors as a priority for future work.

Metabolic associated fatty liver disease (MAFLD) is one of the most prevalent chronic liver diseases worldwide, closely linked to obesity, insulin resistance, and metabolic syndrome [1]. While liver biopsy is the gold standard, it is invasive, costly, and impractical for early detection. Using ex-vivo MRS and gaschromatography (GC-MS), previous studies have demonstrated that the composition of fatty acids in the liver changes as MAFLD progresses from simple steatosis to steatohepatitis. More recently, a study from our group, using MRS signal from the hepatic lipids and principal component analysis identified the 3 main groups (Healthy, steatosis steatohepatitis) by using as input the 7 peaks corresponding only to the fatty acids [2]. This approach shows promise for distinguishing between different stages of MAFLD based on metabolic shifts, providing a potential non-invasive diagnostic method. In this work we want to stablish : if we had access to more peaks (other metabolites/lipids,besides only fatty acids) the classification of MAFLD groups would improve?. Therefore the aim of this pilot study is to answer the question: Is the analysis of more peaks the key to better diagnosis?

The methodology involved analyzing lipids extracted from livers of eNOS KO mice fed a high-fat diet for 0, 4, 8, and 12 weeks (9 males and 10 females). Using a 9.4 T NMR scanner, 24 metabolic peaks—including 7 related to fatty acids—were identified and quantified. Data analysis focused on calculating the area under the curve (AUC) for each peak using NMR TopSpin software with predefined integral files for accuracy. Statistical comparisons between 7 and 24 peaks were performed using ANOVA in PRISM, analyzing data from 19 mice to identify key metabolic markers like olefinic and diallylic metabolites. To interpret the data, Principal Component Analysis (PCA) was applied to reduce dimensionality and highlight key variations across disease stages, improving visualization. K-means clustering grouped samples based on metabolic profile similarities, allowing identification of distinct clusters corresponding to different NAFLD stages.

The analysis showed significant decreases in olefinic and diallylic metabolites after 4 weeks of a high-fat diet(Fig.1), with levels stabilizing at 8 and 12 weeks. These metabolites exhibited an early decline followed by a plateau, highlighting their role as consistent metabolic indicators during disease progression. Other metabolites such as methyl, methylene, alpha, beta, and allylic showed no significant changes, limiting their usefulness for characterizing the disease. PCA and K-means(Fig.2 and Fig.3) clustering revealed distinct groupings in metabolic profiles corresponding to the diet duration, supporting progressive stages of NAFLD. These analyses clearly separated samples by time points, with olefinic and diallylic metabolites as key contributors.

The observed decrease in olefinic and diallylic metabolites is consistent with previous studies[3] reporting significant depletion of long-chain polyunsaturated fatty acids (LCPUFAs), particularly docosahexaenoic acid (DHA) and eicosapentaenoic acid (EPA), in NAFLD/MAFLD patients. These essential fatty acids, known for their multiple double bonds, are crucial components of cellular membranes and precursors for anti-inflammatory mediators. Their reduction likely underlies the diminished olefinic and diallylic signals detected in our MRS analysis, reflecting disrupted lipid metabolism and increased oxidative stress associated with disease progression. The absence of significant variation in other metabolites highlights the value of targeting a refined subset of key metabolic markers to accurately characterize NAFLD. In this context, Principal Component Analysis (PCA) and clustering methods have proven effective for uncovering metabolic progression patterns and distinguishing disease stages by focusing on the most informative metabolites, thereby enhancing the specificity and interpretability of the analysis. Focusing on a reduced set of 7 informative metabolic peaks improved clarity and reduced noise compared to analyzing all 24 peaks, which suffered from overlapping signals and redundancy. While the 24-peak model showed narrower intra-cluster variance—potentially beneficial in scenarios with higher experimental or biological noise—the 7-peak model achieved better separation of clinically relevant groups and yielded more interpretable results under controlled conditions.

Therefore, we conclude that for this experiment, analyzing 7 peaks is sufficient, as analyzing 24 peaks does not provide additional meaningful information beyond what the 7 peaks reveal.

The Dual-Echo Steady-State (DESS) sequence can be used for T2 mapping [1]. This sequence involves low flip-angle radiofrequency pulses, making it an interesting alternative for imaging at high and ultra-high field due to Specific Absorption Rate limitations. The DESS sequence has recently been optimized for brain T2 mapping by mitigating artefacts due to breathing and underlying B0 field variations [2]. Despite the significant improvement and validation against a spin-echo reference in phantoms, the estimated apparent T2 remained biased in highly concentrated agar-based phantoms [2], suggesting magnetization transfer (MT) influences. Furthermore, in brain tissues, water compartmentation represents an additional intrinsic source of variabilities when considering a mono-compartment model to estimate T2. This study aims to combine numerical simulations and MRI acquisitions to understand the behaviors of T2 estimates from the DESS method.

Acquisitions were performed at 3 T (MAGNETOM Prisma, Siemens Healthineers, Erlangen, Germany) on one volunteer to estimate biophysical parameters required for the modeling and evaluate the consistency with numerical simulations. The MRI protocol included a 3D DESS sequence for T2 mapping, complemented with an MP2RAGE for the required T1 prior. The T2 estimation with the optimized DESS protocols was performed as described in [2]. In addition, a quantitative MT (qMT) protocol was acquired following the methodology in [3] for the estimation of the Macromolecular Proton Fraction (MPF) from the binary spin-bath model, used as a surrogate of myelin water pool size (Myelin Water Fraction; MWF). A summary of the sequences’ parameters is provided in Table 1. DESS and qMT frameworks were both corrected for transmit field inhomogeneities. T2 measurements were evaluated in representative regions of interest (ROI) from white (JHU probabilistic atlas [4]) and deep gray (recon-all from FreeSurfer [5]) matter, using the anatomical UNI image from the MP2RAGE acquisition. Analyses were complemented by simulations to explain the experimental measurements. A realistic 4-pool model was used [6], considering two exchanging water reservoirs (namely myelin water [MW] and intra/extra-cellular water [IEW]), each of them exchanging with two bound reservoirs (myelin and non-myelin). Synthetic signals associated with the DESS sequence were generated using the Extended Phase Graph formalism [7], considering biophysical parameters from Manning et al. [6] (Fig. 1), and as a function of the MWF. Synthetic signals, taken as the summation of MW and IEW transverse magnetizations (as experimentally observed), were then fitted to the mono-component T2 model. As a first approximation, MWF and myelin/non-myelin bound pools were assumed of identical size (Fig. 1). To isolate the effect of exchanges and bound proton contribution to the signal build-up, simulations were also run by considering a pure water-only 2-pool model, while accounting or neglecting inter-pool exchanges.

Experimental apparent T2 maps from the DESS configurations and MPF map are shown in Fig. 2, and averaged values from representative ROIs and associated simulation results in Fig. 3. Globally, the TR=30 ms configuration yielded higher apparent T2 values (spanning from 49 ms to 63 ms in WM, and 44 ms to 66 ms in GM) than the TR=15 ms one (42 ms to 53 ms in WM, and 42 ms to 59 ms in GM). Simulation results using the 4-pool model best captures the dynamic as well as variations with TR. Conversely, both water-only 2-pool models fail to predict the behaviors (opposite trend as a function of TR with ΔT2<0 ms for MPF>6.5%; Fig 3c), especially as myelination increases (predicted T2 values > 65 ms for MPF>7.5% overall).

This preliminary work suggests that accounting for MT effects and water compartmentation using a 4-pool model may be essential to explain the apparent T2 estimations using a DESS sequence. In contrast, only accounting for water compartmentation fails to predict the associated trends with sequence parameters, emphasizing the importance of bound reservoirs on the DESS signal build-up. Compensating for these confounding factors, however, would require extensive MT- & multi-compartiments-sensitized acquisitions to solve for a 4-pool model, which is currently unfeasible within clinically acceptable scan times, and calls for a consensus on the protocol design to mitigate variabilities.

This study highlighted the impact of MT (and associated T1 relaxation), water compartmentation and exchanges in apparent T2 measurements estimated by DESS, drawing attention to the limitations of simplified biophysical models for the quantitative description of biological tissues. More volunteers are required so as to refine these results, and further experiments will be performed at 7 T to evaluate the impact of field-dependent variables such as T1 [8].

The purpose of this study was to evaluate different application scenarios of ultrashort echo time (UTE) magnetic resonance imaging for investigating short-T2 components in the Achilles tendon and to assess their impact on the characterization of the tendon’s compositional structure.

A custom time-interleaved 2D-UTE pulse sequence, as illustrated in Fig. 1, was implemented to enable the acquisition of up to 64 echo time images (1). Measurements were performed with and without fat saturation (FatSat). Ten healthy volunteers (age 33.3 ± 6.91 years) underwent UTE MRI of the Achilles tendon. Two different approaches were compared to address fat contamination: (A) a Bicomponent model with FatSat, and (B) a Tricomponent model using a fat separation technique. Additionally, the impact of magnitude versus complex data reconstruction on compositional assessment was evaluated. Using multi-echo 2D-UTE images acquired at 64 echo times ranging from 60 µs to 32 ms, T2* relaxation decays were quantified with the following four data fitting models(2, 3): a) Bicomponent magnitude data model b) Bicomponent complex data model c) Tricomponent magnitude data model d) Tricomponent complex data model All four models yielded amplitudes of the short (AS) and long (AL) T2* relaxation components. Based on the estimated parameters, the short (FS) and long (FL) component fractions, expressed as percentages, were calculated as (4): FS = 100xAS/(AS + AL) and FL = 100xAL/( AS + AL ) respectively. Image acquisition was performed on a clinical whole-body 3T scanner (Magnetom Prisma, Siemens Healthineers, Erlangen, Germany). All experiments used a whole-body RF coil for transmission and a flexible four-channel phased-array coil for reception. Image reconstruction and data processing were conducted offline using custom routines implemented in MATLAB (The MathWorks, Natick, MA, USA). All analyses were performed using R Statistical Software (version 4.4.1). Normality of data was assessed by Shapiro–Wilk test and Q-Q plots. A two-way repeated measures ANOVA was performed with probands as the within-subject identifier and data acquisition/reconstruction type as within-subject factors. Post hoc pairwise comparisons were conducted using estimated marginal means with Holm-adjusted p-values to control for multiple comparisons. In the statistical analyses, the following data acquisition/reconstruction type scenarios (i.e., post-hoc pairs) were considered: a) FatSat/Magnitude vs. non-FatSat/Magnitude b) FatSat/Complex vs. non-FatSat/Complex c) FatSat/Magnitude vs. FatSat/Complex d) non-FatSat/Magnitude vs. non-FatSat/Complex, and P-values < 0.05 were considered statistically significant.

Ten healthy volunteers (3 males, 7 females; mean age: 33.3 ± 6.9 years), with no history of Achilles tendon injury, participated in the study. Institutional approval was obtained, and all participants provided written informed consent prior to the study. Figure 2 shows representative images acquired with and without FatSat, along with data fitting results from all four models applied to a region of interest (ROI) selected in the mid-portion of the Achilles tendon (MID). The results from the ANOVA analyses are summarized in Tab. 1, and the short component ratios Fs calculated from different acquisition/reconstruction scenarios are presented in Fig. 3.

Regardless of the Achilles tendon region, fat saturation, reconstruction type, and their interaction all had a statistically significant effect on the short component ratio (p < 0.05). The interaction between fat saturation and reconstruction type was not of primary interest and had a minimal effect—explaining only 0.6% of the variance in the INS region (p = 0.025), 4.7% in the MID region (p < 0.001), and 3% in the MTJ region (p < 0.001). In all tendon regions, fat saturation increased the short component ratio for both reconstruction types. Conversely, complex reconstruction consistently resulted in a higher short component proportion, with a more pronounced difference observed in the non-FatSat condition. The estimated marginal means and their standard errors were consistent with the differences in group means and their standard errors calculated from the original data.

Both the data acquisition and reconstruction methods significantly impact the short T2* component ratio and should be considered when comparing results across studies. . Acknowledgements This work was funded by Czech Science Foundation grant no. GF25-19984L. Also, this research was funded in whole or in part by the Austrian Science Fund (FWF) 10.55776/PIN5555423. We acknowledge the core facility MAFIL supported by MEYS CR (LM2023050 Czech-BioImaging), part of the Euro-BioImaging (www.eurobioimaging.eu) ALM and Multimodal Imaging Node (Brno, CZ).

Isotropic test fluids or gels with well-defined apparent diffusion coefficients (ADC values) are essential tools for testing and validating diffusion-weighted imaging techniques [1-5]. We have previously shown that aqueous soy lecithin solutions are a beneficial material for the construction of diffusion phantoms with tissue-like ADC values [6]. Soy lecithin added to water provides a wide range of adjustable ADC values without introducing spectral signals. In addition, soy lecithin is compatible with agarose, allowing simultaneous tuning of ADC and T2. However, the behavior of aqueous soy lecithin solutions at different measurement temperatures has not yet been investigated. Temperature is a critical factor, as the molecular mobility of water – and therefore the ADC – is strongly temperature-dependent [5,7]. Consequently, measurements taken under different conditions may yield different ADC values. To address this issue, we systematically investigated the temperature-dependent behavior of aqueous soy lecithin solutions. Solutions with varying soy lecithin concentrations (0-10% (w/v)) were prepared, and for each concentration, ADC values were measured at seven different temperatures ranging from 3°C to 37°C. Based on these results calibration curves were derived that can be used to correct temperature-related ADC variations.

Data acquisition and analysis: MRI was performed on a clinical whole-body 3.0 Tesla scanner (MAGNETOM Prismafit, Siemens Healthcare, Erlangen, Germany) using an 18-channel body-array coil. Measurement temperatures were as follows: 3°C, 10°C, 15°C, 20°C, 25°C, 30°C, and 37°C. Image processing, including mapping and analysis, was performed offline in MATLAB (R2022b, MathWorks, Natick, MA, USA). DW-MRI was performed using a readout-segmented echoplanar imaging sequence with 4 different b-values (0, 50, 500, 1000 s/mm2). TR and TE were set to 5000 ms and 51 ms, respectively. ADC maps were calculated from the acquisitions with multiple b-values using a log-linear fitting of the signal intensities. Materials and sample preparation: Soy lecithin (Carl Roth, Karlsruhe, Germany) was dissolved in deionized water by magnetic stirring for 10 minutes at 650 rpm. The concentrations of soy lecithin varied from 0-10% (w/v). After preparation, the solutions were filled into sterilized polypropylene tubes (Greiner Bio-One, Frickenhausen, Germany) and brought to the desired temperature in a heating cabinet. To keep the temperature stable during the measurement period, the solutions were placed in a thermally insulated polystyrene box for the measurements.

The relationship between ADC and temperature was found to be mono-exponential (Eqs. 1) across all tested soy lecithin concentrations (Figure 1). ADC [T]=a_C ∙ exp(b_T ∙ T) (1) The growth rate b_T of the exponential fit was identical for all concentrations, indicating a uniform temperature behavior for all aqueous soy lecithin solutions (Table 1). In contrast, the intercept a_C varied with concentration, reflecting the concentration-dependent decrease in ADC. This relationship followed a biexponential trend, independent of the measurement temperature (Equation 2, Figure 2). a_C = ADC [C_lec ]=A ∙ exp(-B ∙ C_lec) + C ∙ exp(-D ∙ C_lec) (2) Based on these findings, a combined model was established to describe the ADC as a function of both soy lecithin concentration and temperature: ADC [C_lec,T] = (A ∙ exp(-B ∙ C_lec) + C ∙ exp(-D ∙ C_lec)) ∙ exp(b_T ∙ T) (3) This model accurately described the experimental data (R2 = 0.966) and reliably captured the effects of soy lecithin concentration and measurement temperature on the ADC value (Figure 3).

In this work, calibration curves were derived to describe the temperature and concentration dependence of the ADC value in aqueous soy lecithin solutions. The resulting model provides a practical tool for correcting temperature-related ADC variations and may help to improve comparability of DWI measurements under varying experimental conditions.

Concussion, a mild form of traumatic brain injury (mTBI), is linked to subtle microstructural changes often missed by conventional imaging [1]. In this study, we applied intravoxel incoherent motion (IVIM) MRI to compare perfusion and diffusion between concussion patients and healthy controls (HCs), using both the standard bi-exponential model [2] and an extended tri-exponential model that adds a fast diffusion component, potentially reflecting free water or rapid microvascular flow [3]. The bi-exponential model estimates tissue diffusion (D), perfusion-related diffusion (D*), and perfusion fraction (F), while the tri-exponential model provides separate parameters for tissue (Ds, Fs), perfusion (Dp, Fp), and fast/free diffusion (Df, Ff). We hypothesize that the tri-exponential model may offer greater sensitivity to post-concussion changes and aimed to identify IVIM-derived biomarkers distinguishing concussion from HC.

We recruited 20 HCs (9 F, mean age 29.6±6.5) and 20 individuals with mTBI (10 F, mean age 28.0±7.5). All participants completed the Montreal Cognitive Assessment (MoCA) [4]; the mTBI group also underwent Glasgow Outcome Scale-Extended (GOS-E) testing [5]. Diffusion MRI was acquired on a 3T Philips Ingenia using a modified IVIM protocol with 14 b-values (10–3000 s/mm²) and up to 12 directions per b-value. Preprocessing included denoising and motion, distortion, and eddy-current correction. Bi- and tri-exponential IVIM maps were generated using a custom Python script. Test-retest reliability was assessed in 11 HC (4 F, mean age 31.8±6.6) using ICC [6] and Bland-Altman [7] analyses. Voxel-based analysis was conducted in R, using a linear model adjusted for age, sex, and motion. Group-level statistics included pTFCE [8] and Benjamini-Yekutieli FDR correction [9]. Effect sizes were computed using Hedges' g [10], with α = 0.05 and power = 0.80 (large effect: g>0.99).

Lower MoCA scores were found in the mTBI group (HC=28.5±1.7, mTBI=25.5±3.1; t=4.2, p<0.001), while the mean value of GOS-E for mTBI was 6.40±1.10, indicative of moderate to good recovery. No difference in age was observed (t=0.73, p=0.47). Figure 1 (a) shows ICC distributions across IVIM metrics. Slow diffusion parameters, D (ICC = 0.72±0.22) and Ds (ICC = 0.77±0.18), had high reliability. In contrast, bi-exponential D* and tri-exponential Dp had poor reproducibility. Perfusion fractions showed high reliability for Fp (0.71±0.20) and Fs (0.69±0.21) and moderate reliability for F (ICC = 0.63±0.24) and Ff (0.55±0.23). Figure 1 (b) presents Bland-Altman plots for best/worst BIAS subject (left and right panels, respectively). Metrics with high ICCs (D, Ds, and Fs) had narrow limits of agreement (LOA), indicating strong test-retest agreement. D* and Dp showed high variability and wide LOA. Group comparisons, focusing on metrics demonstrating good reliability (ICC>0.50), revealed significant tri-exponential IVIM changes in the concussion group (Figure 2). Specifically, the slow diffusion coefficient (Ds) was reduced in concussed individuals compared to HCs, particularly within frontal and parietal regions. This reduction may indicate a loss of slow-diffusing components, potentially stemming from microstructural disorganization or diminished hindrance due to tissue swelling. Similar patterns of reduction were observed for the perfusion-related fraction (Fp) and the fast perfusion fraction (Ff). Conversely, the fraction associated with tissue diffusion (Fs) was elevated in the concussion group, suggesting changes in the tissue microenvironment that might enhance the relative contribution of restricted water diffusion, potentially linked to edema or microvascular changes. Bi-exponential IVIM results (Figure 3) showed lower tissue diffusion (D) in concussion subjects relative to controls in parietal/insula/temporal regions; minor D increments were observed but lacked substantial effect sizes. The perfusion fraction (F) was significantly lower in the concussion group, suggesting reduced microvascular perfusion post-concussion. A comprehensive summary of all findings is presented in Figure 4.

Our findings support the utility of IVIM MRI, particularly the triexponential model, in detecting subtle brain changes after concussion. The triexponential metrics showed higher test-retest reliability and revealed more widespread group differences through voxel-based analysis. Reduced Ds in concussion may reflect microstructural disruption or decreased restricted diffusion, while elevated Fs suggests changes in the microvascular or interstitial environment. Additionally, decreased Fp and Ff can be connected to tissue damage, edema, and cell loss. The biexponential results also showed lower F metric in concussion, indicating possible vascular alterations.

Overall, these IVIM-derived metrics, especially from the triexponential model, may serve as promising biomarkers for concussion-related brain changes.

The long-term fate of gadolinium (Gd) from GBCAs in the human body remains a subject of scientific investigation. While these agents have revolutionised diagnostic imaging, the potential for Gd retention in tissues raised concerns worldwide several years ago through unexpected hyperintense signal areas in clinical scans. This study investigates the interactions of Gd with biological molecules, focusing on the role of sugar structures in Gd retention. Using dextran sulphate (DS) as a model compound, we aim to elucidate how different molecular weights of polysaccharides influence Gd behaviour and detectability. Our investigation provides insights into the chelation process between Gd ions and sugar molecules, molecular weight dependent affects Gd relaxivity and potential mechanisms of Gd shielding that may affect MRI detection. This research contributes to the growing field of knowledge on Gd deposition and aims to improve our understanding of the safety of MRI contrast agents. By exploring these fundamental interactions, we are paving the way for more accurate assessment of Gd retention and potentially improved contrast agent designs in the future.

All MR measurements were performed on a 9.4 T MRI system (Bruker, Germany). T1 measurements were performed using a dephasing recovery sequence consisting of 50 π/2 pulses with subsequent gradient spoiling and image acquisition. R1 values were calculated from ROI-averaged values from R1 maps. To model the transchelation processes of Gd-ions to different sugars, DS (8 kDa – 500 kDa) in combination with GdCl3 were used as a model system. Light scattering measurements were used to investigate the increase of the aggregate size as a result of Gd-ion binding to DS and CSA.

An increase in R1 was observed after the binding of Gd-ions to the added polysaccharides. Subsequently, a reduction of R1 could be observed with increasing polysaccharide concentration. Figure 1 shows R1 as a function of the concentration ratio between DS with different MW (with adjusted to total number of sugar units) and Gd-ions in solution. In all measurements, R1 increases from 0.6 s-1 (R1 of 25 μM Gd3+ in H2O) to ~0.8 s-1 and to ~0.85 s-1 for (DS) of MW = 8 kDa, 20 kDa and 500 kDa, respectively. The maxima are reached at ratios of ~101.1, ~101.3, and ~102.3. Subsequently, R1 decreases to ca. 0.4 s-1, 0.45 s-1, and 0.5 s-1 at ratios of ~103.1, ~104, and ~105 for (DS) of MW = 5, 20, and 500 kDa, respectively.

Glycosaminoglycans (GAGs) display a heterogeneous distribution within tissues, thereby creating local microenvironments with the potential to act as hotspots for Gd deposition. The results obtained at first support the hypothesis that Gd binding to sugar structures, increases longitudinal relaxation rates (R1) thus providing a plausible explanation for clinical observations of tissue hyperintensities. This phenomenon can be attributed to the lower tumbling rate of Gd-sugar complexes in comparison to free Gd ions, resulting in enhanced relaxivity at low [GAG]:[GdCl3] ratios. However, at higher sugar concentrations, this initial increase in R1 is counteracted by a loss of signal. This shielding effect is attributed to sulphate groups occupying Gd coordination sites, thereby limiting water access to the paramagnetic center. At the same time, Gd-mediated cross-linking of the sugar chains, as shown by light scattering data, induced aggregation and amplifies particle size. It is noteworthy that DS structures with lower molecular weight (MW) accelerate this masking phenomenon, presumably due to steric constraints that promote tighter coordination and faster saturation of binding sites. This MW-dependent shielding may lead to a potential underestimation of uncleared Gd3+ in clinical MRI, as aggregated complexes may avoid detection despite persistent tissue deposition. These findings underscore the necessity to refine imaging protocols and analytical methods to account for polysaccharide-Gd interactions, particularly in tissues with high GAG concentrations. It is recommended that future studies investigate how variations in GAG composition and sulphation patterns affect Gd retention and visibility, thereby addressing a critical gap in understanding long-term deposition.

Our results show that the binding of Gd ions to polysaccharides and the associated increase or decrease in R1 is strongly dependent on the molecular weight (MW) in addition to the concentration ratios. Using model structures such as DS, we have shown that in relaxometry experiments, in addition to hyperintense signals, a clear masking of the Gd ions can be caused by the polysaccharide concentration. We were able to demonstrate this effect for different MWs and showed that the signal gain and loss is characteristic for each individual molecular size. This highlights the need for further research in this area to ensure that the problem of deposited Gd ions in biological tissues is not underestimated in the future.

Phantoms, as reference tools for quantitative MRI, maintain data reproducibility by ensuring scanners correct operation and by leading to improvements in image quality. They allow development and validation of new methods and comparison between acquisition and analysis performances[1–3]. Diffusion phantoms are more commonly developed consider isotropic diffusion[4–9]. Recent literature shows only a few examples of anisotropic diffusion phantoms in more sophisticated and updated versions[10–12], considering in few cases also different fiber crossing angles or a different fiber density rate[13]. Furthermore, few recommendations are available for preclinical purposes[2] and thus less studies focus on diffusion sequence optimization, giving a reference for many possible ex-vivo/in-vivo conditions. The aim of this preliminary study is to optimize a DTI-EPI sequence, exploiting a preclinical diffusion phantom, considering DTI metrics, image quality in terms of SNR maximization and tractography quality as target of optimization.

An MRI diffusion phantom produced in a preclinical custom-made size was used (Phantom Metrics, Psychology Software Tools, Inc., Pittsburgh, PA, USA). The phantom consisted of a 4mm x 4mm x 20mm nylon fiber bundle (Taxon™ technology[14]). Phantom manufacture: 1,470 round fibers (diameter range = 105–110μm, 6,532 water filled Taxon holes (inside diameters = 0.68-0.78μm) per fiber, hindered water spaces between the round fibers estimated as 8% of the fiber bundle total volume. The phantom was placed in a water filled 50ml falcon tube, near the tube surface. Diffusion imaging was performed on a Mediso nanoScan®7T MRI scanner with a DTI-EPI sequence. Sequence parameters were tuned on the phantom, keeping a resolution of about 200μm, 25 diffusion directions and b=500s/mm2. Main changes were focused on single/multi-shot choice, TE/diffusion time trade off, imaging and diffusion gradient parameters as well as encoding direction to minimize artifacts and number of excitations. Image Signal to Noise ratio (SNR) maximization has been considered first target of sequence optimization. Data processing was performed through the ExploreDTI tool[15]. After data reconstruction from raw diffusion-weighted images and LPCA Denoising Algorithm application[16], ADC, FA, MD and AD maps were calculated. The known fiber density and direction and thus the diffusion behavior have been taken as reference to select the best parameters setting, while further considerations have been driven about image and tractography quality.

The first step of parameters optimization resulted in the following set: amplifier configuration = 750V, 4 shots, 80% symmetric Partial Fourier (PF) acquisition (0.22x0.28mm2) in the phase encoding direction, diffusion gradient ramp time = 1.016ms, duration(δ)=1.2ms, and interval(Δ)=9ms. With respect to other considered parameter settings, this configuration became the reference since it gave the best results in terms of FA (0.72), as precise representation of the phantom anisotropic nature (tractographic reconstruction shown in Fig.1). DTI metrics obtained in function of parameters variation are reported in Table 1. Starting from the reference setting, image and tractography quality have been quantified using number of reconstructed tracts (Fig.2) and SNR calculation (Fig.3) respectively. Both indices have been considered in function of number of excitations (NEX 2 to 5), number of shots (1,2,4), two PF types and when applying or not the denoising algorithm.

The reference setting resulted to be the best parameters configuration, with metrics comparable to literature values[17]. The higher FA together with the lower AD confirm a high diffusion along the known direction of the fibers, more than in the axial one. Using a different amplifier configuration and adjusting the diffusion gradient parameters, high FA is obtained when ramp time and duration are similar and when applying the symmetric 80%PF in case of 4 shot and asymmetric in case of 2 shots. A sensible FA reduction can be noticed especially when using a 1 shot. Indeed, even if the acquisition time decreases when using 1 shot and could be better in terms of in-vivo applications, the echo time considerably increases, producing less signal from the phantom and therefore low SNR. Increasing NEX confirmed an increase of SNR values, with a more evident trend when considering the non-denoised images, even if for these latter, absolute values 10 times lower than the corresponding ones in denoised images can be noted. The same clear trend appears also in tractography quality for the non-denoised images, while a better reconstruction in terms can be achieved even for the 1 shot but applying the denoising algorithm. No general substantial differences are produced between the 2 PF types.

In this preliminary study, a reference optimized sequence has been obtained, with improvements in term of DTI metrics, after acquisitions with different parameters sets.

MRI is capable of generating a variety of contrast mechanisms. The acquired signal used to generate the image depends on a range of parameters, where each of them is either object specific or method specific. Knowledge of this dependency provides the possibility to adjust the signal intensity of any spin isochromat by modifying method specific parameters in the dependency. As this reliance is known for a sequence that is employed to generate the image, the signal intensity of and therefore the contrast between different isochromats in MR images are adjustable. In this work we present proof of this concept to UHF-MRI (7 T) aiming for applying it to knee imaging.

The signal generated by Spin-Echo (SE) sequence [1] was simulated using in-house MATLAB [2] scripts based on closed-form solution of the Bloch equations for SE signal. Considering a specific range of TRs and TEs given to the simulation while other parameters are known, the outcome is the range of TRs and TEs at which a certain level of contrast between two spin isochromats will be generated. Attaining the desired level of contrast was verified on a phantom. The phantom consisted of an apple, a potato and a citrus. Images were acquired using SIEMENS 7T MR Imager (MAGNETOM Terra) and dedicated knee coil (1Tx 28Rx Knee Coil 7T Clinic). Due to images of each of the three objects being not isointense, relatively homogeneous regions were selected for relaxation times calculation. Using inversion recovery acquisition method (TIs=40, 100, 300, 600, 1000, 1500, 3000, 5000 ms, TRs=TI+10000 ms) and an exponential fitting to the acquired data, we estimated the T1 value of apple and potato in a relatively homogeneous region (7 slices (apple), 19 slices (potato), 30 slices (total),2.5/0.5 mm thickness/gap). SE acquisition method applied at multiple TEs (TEs=6, 8, 12, 15, 30, 45, 65, 95, 120, 150, 180 ms, TR=7500 ms) and an exponential fitting were performed to measure T2 of the two objects in almost the same regions (8 slices (apple),18 slices (potato), 40 slices (total), 2.5/0.5 mm thickness/gap). SE acquisition with very short TE and very long TR (TE=6 ms, TR=15000 ms ) to leave the signal weighted only by proton density was used where relative signal intensities of the two objects from regions of equal size delivered relative spin density which was then averaged across selected slices. The contrast between apple and potato was calculated in two slices (2.0/0.1mm thickness/gap) which were adjacent and among the slices where relaxation times were calculated. Our evaluation of contrast adjustment was performed in two sets of SE imaging, one with long TR=5580 ms (with five contrast levels from 95±5% down to 15±5%) and the other with relatively shorter TR =1780 ms (with four contrast levels from 95±5% down to 65±5%). By feeding the simulation with the relaxation times, spin density ratios of the two objects, and the TR and the flip angle of the sequence refocusing pulse, we calculated TEs at which the contrast lies within the desired level and range. The range of TEs given to the simulation was 10 to 120 ms with resolution of 50 points equally spaced. These TEs were given to the sequence to generate images with adjusted contrast.

Figure 1 shows the result of simulation for a specific TR of 5580 ms. For each level and range of contrast the calculated TEs are presented in the corresponding plot. The measured relaxation times of the two objects and their relative proton density ratios normalized to that of apple are given in Table 1. Tables 2 and 3 manifest the desired levels of contrast, the TEs that were calculated and entered into the SE sequence for each level, the measured CNR and its level relative to the reference (ref) for two sets of SE image acquisitions with different TRs. The levels are given in percentage relative to the maximum achievable contrast. Measured mean CNRs lie within the range of 10% which was set for each level of contrast.

This work is an initial step towards generating contrast between tissues at any desired level in MR images, e.g. in knee MR images generated at UHF (Ultrahigh Field). To achieve this, all parameters that affect the MR signal intensity must be considered in the whole process of contrast adjustment. The type and number of some of these parameters are dependent on the sequence which is employed to generate the images. The accuracy of the calculations could also be affected by the applied reconstruction technique e.g. when parallel imaging is employed to accelerate the acquisition. Knowledge of precise values of relaxation times of the spin isochromats that are imaged elevates the accuracy of the contrast adjustment.

In this work we present a proof of concept of the possibility to gain specific levels of contrast between different objects in MR images with the support of simulation. The approach could be applied to a variety of sequences and paves the way for establishment of a fully automated contrast adjustment for MR images.

Chemical Exchange Saturation Transfer (CEST) MRI is a promising molecular imaging tool based on proton exchange between the target molecule and water. Because of its unique contrast mechanism, the interest for this technique is still increasing for both clinical and preclinical applications. Many molecules can act as CEST agents, both endogenous molecules and exogenous agents. APT (Amide Proton Transfer) CEST is one of the most widely used CEST techniques. APT signals primarily arise from the proteins and peptides found in tissue[1]. The technique is most commonly used for detecting cancer[2, 3] and ischemic stroke[4]. CEST MRI of iopamidol offers a noninvasive method for assessing tissue pH by leveraging the pH-sensitive properties of iopamidol contrast agent[5–7]. This approach provides valuable diagnostic insights, especially for detecting pathological conditions such as tumors[8]. Our aim was to investigate the performance of CEST-SSFP-FID[9] sequence in terms of APT and iopamidol CEST using phantoms of poly-L-lysine (0.1% w/v) and iopamidol (30mM), respectively.

All measurements were performed on a nanoScan® 7T MRI Mediso scanner. For each CEST-SSFP-FID acquisition, the following parameters were used: TE=1.85ms, TR=3.71ms, matrix = 64 × 64, FOV = 38.4 mm x 38.4 mm, slice thickness = 2mm (single-slice), flip angle=30°, NEX=1. Centric k-space filling was applied with 10 dummy scans. Presaturation delay of 5s was set between the readout and the following CEST saturation pulse. The total scan time was 2min 37s to acquire CEST images at 28 saturation frequencies. Three CEST saturation conditions have been investigated. The use of different number of pulses (single or train of 10 hard pulses with interpulse delay of 10µs) during the same total saturation time. Then, the effect of saturation power (B1). Lastly, the application of a spoiler gradient after the saturation. We acquired 28 saturation frequencies for both substances, with iopamidol ranging from -8ppm to +8ppm and poly-L-lysine ranging from -6ppm to +6ppm. For both phantoms, less points have been acquired in the negative range, since the target molecules have no influence on water signal in that range. A reference image was obtained with saturation frequency at -150ppm, where no CEST effect was expected. Z-spectra were obtained from a circular ROI placed in the center of the phantom and avoiding artifacts at air-phantom border. Image intensities were averaged within this ROI and expressed as a percentage of the ROI-averaged values obtained from the reference image. These values have been used to generate the Z-spectra and to calculate the asymmetric magnetization transfer ratio (MTRasym)[10] for iopamidol at 4.2ppm and 5.6ppm and for poly-L-lysine between 2 and 4 ppm.

Figures 1 and 2 show Z-spectra obtained with different saturation powers for iopamidol and poly-L-lysine, respectively. Higher contrast peaks are expected at 4.2ppm and 5.6ppm for iopamidol and at an extended area between 2 and 4 ppm for poly-L-lysine. Figure 3a and 3b show Z-spectra obtained using 3µT saturation power for both iopamidol and poly-L-lysine. Figure 3a shows the effect of switching on and off the spoiler gradient after CEST saturation pulse, obtaining no evident difference in the Z-spectra and MTRasym values for both substances. Figure 3b focused on the effect of pulse train compared to single pulse.

As expected, higher loss of signal on Z-spectra as well as higher MTRasym values (as shown in Figure 4) correspond to power value increase. However, the peak width also increases and this could make the target effect less detectable. For CEST contrast of iopamidol, 3µT can be a good trade-off, since it produces high contrast at 4.2ppm and 5.6ppm, keeping the water peak sufficiently narrow. Additionally, the OH- group appears more evident with 3µT compared to 6µT. For CEST contrast of poly-L-lysine, 3uT is also a good trade-off. No substantial difference can be noticed in Z-spectra and MTRasym values for both substances, demonstrating the high quality of the pulse train produced by this instrument. Typically, the maximum contrast for APT CEST happens at 3.5ppm, but acquiring poly-L-lysine on this scanner the signal loss in Z-spectra and MTRasym values result higher at 2.5ppm for each saturation conditions.

CEST-SSFP-FID sequence is a fast and reliable method to measure the CEST effect of APT or iopamidol. The CEST application package provided by Mediso on the nanoScan® 7T scanner was validated on phantom scans and the CEST saturation pulse could be optimized by using poly-L-lysine and iopamidol, so in-vivo studies can be carried out in the future.

Hypoxia is a cancer biomarker indicating tumor aggressiveness. Incentive exists to use non-invasive MRI technology to map and quantitatively measure hypoxia. However, traditional in-vitro hypoxic models are incompatible with the MRI experiments that are needed to develop MRI-based hypoxia quantification. Microfluidics technology creates size-controlled 3D cell spheroids wherein the level of hypoxic condition varies with the spheroid size. Using hypoxic spheroids for MRI-based determination of hypoxia has not been investigated. The objective of this research was to demonstrate the viability of using hypoxic cancer spheroids in microfluidic chips to quantify hypoxia using MRI.

Microfluidic chips were utilized to seed, form and culture human colon cancer (HCT116) cells using techniques previously established (Refet-Mollof et al, 2021). A total of 120 hypoxic spheroids (~ 700 µm diameter) and 120 normoxic spheroids (~ 250 µm diameter) were formed; the size of the spheroids was verified with microscopy. The proportion of silicon to media or cells in the chips was constant. A western blot (the gold standard of hypoxia quantification) was performed to quantify the hypoxia-inducible carbonic anhydrase IX (CAIX) protein in half of the normoxic and half of the hypoxic spheroids, verifying the hypoxic content prior to the experiment, and also demonstrating the level of hypoxia in the hypoxic versus normoxic cells. The chips with remaining spheroids were scanned in a 3 Tesla Siemens MRI machine. AT2-weighted sequence (resolution of 0.8 x 0.8 x 0.7 mm3) was employed. Medical Image Merge (MIM, version 7.2.8) was used to measure the signal from each spheroid in the microfluidic chip from a 1 mm diameter region-of-interest. The signals from both normoxic and hypoxic spheroids were separately averaged; standard deviations were calculated. A Student’s t-test was applied to assess significant differences in signals between the normoxic and hypoxic spheroids. After the scans, the CAIX levels in the scanned chips were measured with western blot to verify that hypoxia was maintained throughout the MRI scan.

Individual wells (~ 1 mm) were resolved on the MRI images, indicating the feasibility of using microfluidic chips for cellular studies with MRI. Figure 1 displays an MRI image of the hypoxic spheroids in the wells, illustrating their resolution on MRI. The hypoxic spheroids yielded significantly less signal compared to the normoxic spheroids (normoxic: 1530 +/- 294, hypoxic: 1425 +/- 269, p < 0.05). The loss in signal is expected given that the hypoxic condition decreases T2 relaxation times. Western Blot results confirmed the hypoxic condition during the MRI scans and that the hypoxic spheroids had more CAIX protein than the normoxic spheroids.

This research demonstrates the feasibility of using microfluidic chips for in-vitro studies of hypoxic cells with MRI, as well as other cellular studies. Further studies will explore using additional MRI sequences, quantifying various levels of hypoxia, and developing complex materials using microfluidic chip technology. Eventually, the research will be expanded to cellular models that more closely resemble human anatomical systems, while maintaining the ability to quantitatively study hypoxia.

A T2-weighted MRI sequence yielded significantly less signal from hypoxic spheroids versus normoxic spheroids in microfluidic chip wells, unlocking the potential for using microfluidic chips in cellular studies of hypoxia and other conditions.

The ethical and sustainable optimisation of animal preclinical studies has becoming an increasingly important matter for enhacing the planning and translational validity of in vivo experiments. The development of in silico realistic anatomic plays a crucial role in addressing this issue. In this study, we present a high-resolution, full-body 3D mouse model designed from magnetic resonance imaging (MRI) scans, to support simulations of non-invasive trans-spinal direct current stimulation (tsDCS) in mice.

Acquisition of MRI data: MRI data were acquired from an ex-vivo male C57BL/6 mouse (sacrificed under the Regierungspraesidium Tübingen animal experimentation license no 1522) at postnatal day 30 (PND 30), using an ultra-high-field 11.7 Tesla small animal MRI system (BioSpec 117/16, Bruker BioSpin, Ettlingen, Germany). To visualize tissue interfaces and structures with susceptibility contrast, FLASH T2* acquisition protocol was applied: TR/TE = 1800/4.25 ms, flip angle = 25.0°, matrix size = 250 × 230, spatial resolution = 100 × 100 × 350 µm³, bandwidth = 65.8 kHz, and 150 signal averages, yielding a total imaging duration of approximately 17 hours and 15 minutes. A total of 165 axial slices were acquired. An example of a resulting MRI axial slice is presented on Figure 1. This contrast allowed a visible distinction of key anatomical structures of our model, specifically the following spinal regions: grey matter (GM), white matter (WM), cerebrospinal fluid (CSF), dura mater, intervertebral discs, and vertebrae. Design of the in vivo mouse model: Segmentation of relevant anatomical structures (skin, spinal GM and WM, intervertebral discs, vertebrae, skull and brain, and internal organs) was performed semi-automatically using ITK-SNAP (www.itksnap.org) threshold-based method [1]. For the spinal GM and WM, vertebrae, and visceral organs, manual regions of interest (ROIs) were defined at each relevant region along the cervical, thoracic, and lumbar segments of the spinal cord. The segmented data were converted into surface meshes (.STL file format). Tissue masks were converted into 3D surface meshes and imported into Blender 4.0 (www.blender.org/) and 3-MATIC module from MIMICS software (v16) (www.materialise.com/en/industrial/software/3-matic), for manual refinement, including smoothing, remeshing, and boolean operations to prevent intersections. Additional anatomical structures, such as CSF, dura mater, subcutaneous fat, and muscle, were modelled using defined offsets based on MRI measurements and literature values [2]. All final surface meshes are shown in Figure 2.

The resulting model offers full-body anatomical coverage with a focus on the spinal region, allowing for detailed spatial characterisation of electric field (EF) and direct current (DC) distribution under multiple tsDCS electrode configurations. This virtual platform enables the simulation of stimulation parameters and biological tissue electric properties, facilitating parameter optimisation prior to in vivo experimentation, ensuring both the safety and efficacy of tsDCS protocols [3–6].

To our knowledge, this is the first high-resolution, full-body 3D model of a small animal to be applied in the context of non-invasive spinal cord stimulation. Previous studies have predominantly focused either on larger animal models, such as the rat or cat [5,6], or have been limited to invasive stimulation techniques [7–9]. The need for a full-body anatomical model is particularly relevant, as it allows simulation of multiple electrode montages and current pathways, accounting for other anatomical features that influence EF distribution on the spinal cord. This is essential for accurate prediction and optimisation of tsDCS protocols. Our model was integrated in the first in silico-in vivo study to predict spatial distribution and geometric constraints of the EF resulting induced by tsDCS in SOD1 mice, the established animal model of Amyotrophic Lateral Sclerosis, anatomically similar to the C57BL/6 mice.

This work provides a foundational tool for the translational modelling of neuromodulation in small animal models, bridging preclinical and clinical research. This computational approach contributes significantly to the 3Rs (Replacement, Reduction, and Refinement) in animal research. Future developments further validation with in vivo electrophysiological data, and full pipeline automatization, spanning MRI segmentation, 3D anatomical modelling, and FEM-based simulation, enabling reproducible and scalable generation of subject-specific models.

Magnetic resonance imaging (MRI) plays a central role in preclinical neuroscience by enabling non-invasive visualization of structural and functional changes in animal models [1–2]. High-resolution images with elevated signal-to-noise ratio (SNR) allow precise anatomical assessments and support longitudinal studies [3]. However, challenges such as motion artifacts, field inhomogeneities, and prolonged acquisition under anesthesia persist—even at ultra-high field strengths like 7 T [1,4]. Image noise, in particular, can obscure subtle anatomical features and compromise measurement accuracy, making effective denoising essential for reliable analysis. Traditional methods, including Gaussian filtering and median smoothing, often suppress noise at the expense of structural detail. More advanced techniques like Non-Local Means (NLM) and BM3D improve performance by exploiting image redundancy but are computationally intensive and sensitive to parameter settings [5,6]. Recent advances in deep learning have demonstrated superior denoising capabilities by learning noise patterns directly from data [6]. Convolutional autoencoders (CAEs) reduce noise by encoding images into latent representations that preserve key anatomical structures and reconstruct clean images using pixel-wise loss optimization [7]. In this study, we evaluate the performance of the autoencoder architecture in denoising T2-weighted MR images of rat brains, aiming to enhance image quality while maintaining anatomical accuracy.

16 female Wistar rats were imaged using a 7T preclinical MRI system (MR Solutions Ltd., Guildford, UK). Animals were anesthetized with 1–2% isoflurane and immobilized using a stereotaxic head holder with ear and bite bars to minimize motion. Physiological monitoring included respiration and body temperature (Model 1030, SA Instruments, NY, USA). The imaging protocol included T1-weighted (FSE; TR = 1000 ms, TE = 11 ms, resolution: 0.13 × 0.13 × 0.8 mm³) and T2-weighted (FSE; TR = 2500 ms, TE = 40 ms, resolution: 0.14 × 0.27 × 1 mm³) acquisitions. Clean images were treated as ground truth. Artificial noise—Gaussian, Rician, Rayleigh—was added to generate paired noisy datasets. The dataset was randomly split into training (70%), validation (15%), and test (15%) sets. Data augmentation was applied, including random flipping and rotation. All images were pre-processed through filtering, resizing, and normalization to ensure consistency. A convolutional autoencoder deep learning model was implemented. The model was trained using paired noisy-clean data. Model performance was evaluated using loss function, peak-signal-to-noise-ratio (PSNR), structural similarity index (SSIM), contrast-to-noise-ratio (CNR), and mean squared error (MSE).

In this preliminary study, a convolutional autoencoder was trained to denoise artificially corrupted T2-weighted brain MR images from a dataset of 16 rats. The noisy input images had an average PSNR of 22.40 dB, SSIM of 0.4653, MSE of 0.006571, and CNR of 2.7138. Following 50 epochs of training, the model was evaluated on a separate test set, achieving an average PSNR of 28.50 dB, SSIM of 0.7273, MSE of 0.001435, and CNR of 3.5887. The evolution of validation performance during training is summarized in Table 1. Consistent improvements in PSNR, SSIM, and CNR, along with a decrease in MSE, were observed throughout the learning process. Figure 1 presents the training and validation loss curves, while Figure 2 shows the progression of PSNR values over epochs.

This study demonstrates the feasibility of using convolutional autoencoders (CAEs) for denoising T2-weighted rat brain MRI. The model yielded consistent improvements in PSNR, SSIM, and CNR, indicating effective noise reduction with preserved structural detail. These findings align with recent work highlighting the advantages of deep learning-based denoising over traditional filtering methods in preclinical neuroimaging [4,8]. One limitation of this study is the use of JPEG-formatted images, which may introduce compression artifacts and loss of structural fidelity. Additionally, the model was trained on synthetically noised images without a true ground truth. To address this, future studies should consider phantom-based validation to generate reference data with known signal and noise properties, enabling more accurate and reproducible assessment of denoising performance.Further research is needed to validate performance on raw scanner data and to compare CAEs with alternative deep learning architectures such as GANs.

This study shows that convolutional autoencoders (CAEs) can effectively improve image quality in T2-weighted small animal brain MRI under synthetically induced noise. The model achieved notable gains in PSNR, SSIM, and CNR, demonstrating its capacity to suppress noise.

The orientation of nerve fibers relative to the main magnetic field (B0) of the MRI system has been shown to significantly affect the apparent transverse relaxation rate R2* [1–4]. Many existing studies thereby rely on formalin-fixed brain tissue [5,6]. However, fixation is known to substantially alter relaxation parameters [7,8]. In this study, we investigated the effect of formalin fixation on the orientation dependency of R2*.

MRI brain scans were performed on a Siemens MAGNETOM Prisma 3T system with a 20-channel head and neck coil (Siemens Healthineers, Erlangen Germany). The dataset comprised 22 individuals and was categorized into three acquisition groups: in-vivo, postmortem (PM) in-situ (brain not excised), and PM formalin-fixed ex-situ (brain excised). The in-vivo group consisted of four healthy volunteers (age 29.5 ± 5.6 y; 4 f). The PM in situ group included eleven neurologically healthy deceased individuals (age at time of death 59.4 ± 15.2 y; 4 f, 8 m), as well as six deceased patients with ALS (age 63.2 ± 8.2 y; 6 m). The brains of the ALS patients were excised the day after the in-situ scan and fixed in formalin containing 4.5% formaldehyde (acid-free, phosphate-buffered, Roti-Histofix 4.5%, Carl Roth, Karlsruhe, Germany) at room temperature for 3 months for a subsequent PM ex-situ MRI scan. For the PM ex-situ MRI examination, the excised brains were placed in a spherical acrylic container. To prevent deformation and movement of the brain, custom-made brain-specific plates were 3D printed [7]. After placement of the brains in the containers, they were filled with a proton-free and tissue susceptibility-matched fluid (Galden SV 80, Solvay Specialty Polymers, Tavaux, France) to improve ex-situ MRI scanning conditions [9]. MRI sequences included MP2RAGE or TSE (brain tissue segmentation) [10], DTI with b = 2000 s/mm2, 3 b = 0 s/mm2, N = 64 isotropic (calculation of nerve fiber angle) and multi-echo GRE with 12 TEs, TE1= 5.79 ms, TR = 68 ms (R2* calculation). Data analysis was performed using the FMRIB Software Library (FSL v6.0.7.13) [11–13]. Skull stripping was performed using HD-BET [14]. After that, FSL’s FAST was used to create the white matter mask [15]. DTI data were corrected for eddy currents and head motion with FSL’s eddy_correct. Diffusion parameters were computed with FSL’s DTIFIT. The principal eigenvector of the diffusion tensor was used to calculate the nerve fiber angle. R2* was computed via a voxel-wise two-parameter mono-exponential single decay fit in MATLAB (The MathWorks, Inc., Natick, MA, United States), see equation (1) [3,4,16–19]. R_2,1^*(θ)=a_0+a_1*cos⁡(θ)^2+a_2*cos(θ)^4 (1) Fiber orientations were divided into 18 intervals of 5° each, covering the range from 0° (parallel to B0) to 90° (perpendicular to B0). Additionally, the data across subjects and groups were normalized to ensure comparability. All data underwent a linear registration to ensure the same pixel dimensions and orientation.

Figure 1 displays the normalized R2* values, while Figure 2 presents the absolute R2* values as a function of the fiber angle for the three examined groups. Each graph represents the group’s mean values and standard deviations. The in-vivo group exhibited the most pronounced orientation-dependent variation with normalized R2* values ranging from 0.89 - 1.12. In the in-situ group, relative values ranged from 0.91 - 1.05, while relative values ranged from 0.94 - 1.01 for ex-situ cases. As expected, absolute R2* values were highest in the ex-situ group and lowest in the in-vivo group.

Our results confirm the orientation dependency of R2* in all investigated conditions: in-vivo, in-situ and ex-situ. The strongest effects were observed in the in-vivo group, with deviations exceeding ± 10 %. In-situ and ex-situ conditions showed smaller effects. The three examined groups were scanned at different temperatures: in-vivo at the highest, ex-situ at intermediate room temperature and in-situ at the lowest temperatures. Since the in-situ condition resulted in an intermediate range of R2* values, it is unlikely that the observed differences in orientation dependency between groups are primarily due to temperature. This supports the hypothesis that formalin fixation influenced the observed R2* orientation dependency by altering tissue microstructure through protein cross-linking and dehydration, thereby reducing magnetic susceptibility differences and local field inhomogeneities [20–22]. Additionally, it is unlikely that the observed effect is attributable to ALS-related changes, as no differences were detected in the in-situ orientation dependency between ALS and neurologically healthy individuals.

Our findings demonstrate an orientation sensitivity of R2*, with the lowest orientation dependency observed ex-situ. This suggests that formalin fixation alters tissue microstructure and reduces R2* orientation dependency, emphasizing the need for caution when interpreting ex-situ data from fixed brain tissue.

Neurodegeneration with brain iron accumulation (NBIA) is a group of extremely rare diseases characterized by intracerebral iron accumulation and progressive neurodegeneration. To date, 11 genes have been implicated in NBIA [1]. The pathophysiology of NBIA is partially understood, although several pathways have already been identified, including mitochondrial homeostasis, lipid peroxidation, autophagy and iron metabolism. Although some imaging patterns suggest a specific NBIA subtype, there is no correlation between evolution of iron accumulation, stage of the disease and specific NBIA subtype. A quantitative measure of the iron load could then be used as a biomarker of the disease and for evaluating the pathology severity. Quantitative Susceptibility Mapping (QSM) provides an iron-load quantitative metric [2,3], with promising investigations in NBIA at ultra-high magnetic field [4,5]. Here we show feasibility in a clinical context at 3T.

MRI data were acquired at Bordeaux Hospital (France) on a General Electric 3 Tesla scanner with 48Rx head coil. A 3D gradient echo multi-echo sequence (SWAN) was applied following the consensus guidelines [6] (TA=4.8 min, FoV=224x224x176 mm, voxel size=1 mm isotropic, TR=34.7 ms, TE1=4.34 ms, ∆TE=5 ms, number of echoes=6, flip angle=16°, bandwidth=280 Hz/px, with compress-sensing) after a quality-control analysis on phantom and on a volunteer, the protocol was applied to NBIA patients. QSM reconstructions were performed in a secured-cloud European-based infrastructure[7]. A pipeline was implemented in Python, to compute R2* and QSM maps using MEDI [3,8], and extract values for several regions of interest (ROIs). Basal ganglia ROIs were segmented manually for each subject. Mean and standard deviations were computed in each ROI.

We successfully implemented and validated an acquisition protocol for QSM in accordance with the QSM consensus guidelines, including the ability to use the compressed sensing option to reduce acquisition time without affecting phase for QSM on a clinical 3T system. Based on our quality-control procedures, with the applied protocol, we obtained on phantom a signal-to-noise ratio close to 225 with 1 mm isotropic acquisition providing QSM without significant bias (less than 1 ppb) and a standard deviation below 1 ppb. On the volunteer, average signal-to-noise ratio on the brain was larger than 100, indicating a field-mapping precision better than 1 ppb. Patients analysed are three women, respectively affected by BPAN (linked to a mutation in the WDR45 gene), PKAN (PANK2), and MPAN (C19orf12). Increased iron accumulation in the basal ganglia was observed in NBIA, with a ~5-fold increase in pallidum on average, with values reaching up to more than 1000 ppb (Fig. 1), and a ~2-fold increase in caudate, putamen and substantia nigra on average over each ROIs. Interestingly, our results suggest a specific iron distribution profile in basal ganglia for each patient (Fig. 2), mainly in subtantia nigra and in pallidum.

QSM give access to brain iron accumulation in NBIA patients on a clinical 3T system with clear individual iron accumulation profiles. These preliminary results indicate that QSM could be a useful biomarker to better understand the links between subtypes, affected regions, disease progression and iron accumulation. In the care pathway, it could be used for diagnosis and follow-up in rare diseases involving brain iron accumulation. QSM may provide a useful biomarker to evaluate efficiency and safety in the context of new therapy development.

Automatic brain segmentation offers a scalable and objective alternative to manual methods, enabling the extraction of neuroanatomical features that may otherwise be impractical to assess. Given that patients with Parkinson’s disease (PDP) and healthy controls (HC) can be distinguished with MRI scans [1,2], we explore whether quantitative metrics derived from relevant regions of interest (ROIs) can effectively classify these two groups in an ongoing PD study [3].

45 subjects were included in this study; 24 patients already diagnosed with Parkinson’s disease (13 male, 11 female, mean age 65.8 years (SD=6.7)), and 21 age-matched healthy controls (10 male, 11 female, mean age 60.2 (SD=9.2)). All MRI data were acquired using a Siemens Magnetom Terra 7T MR system (Siemens Healthineers, Erlangen, Germany) equipped with a 1TX/32RXhead coil (Nova Medical Inc, Wilmington, Delaware). The protocol included MP2RAGE [4] and a multi-echo GRE for QSM (ASPIRE sequence [5]). The resolution for both images was (0.75mm)³. TGV was used for QSM-processing [6]. ROIs that are important for Parkinson’s disease do not all have contrast on T1w images, so we used two pipelines for obtaining ROIs. For putamen, caudate, thalamus, hippocampus, amygdala and lateral ventricles we used FastSurferVINN (FSV) [7,8,9]. For substantia nigra (SN), red nucleus (RN) and subthalamic nucleus (STN) we used nnUNet [10] to train a UNet that segments on QSM. For FSV, we first removed the background noise outside the skull of the MP2RAGE UNI image. We then ran FSV on it. The pipeline is illustrated in Figure 1, and Figure 2 shows an example of the segmentation. The UNet was trained on 15 manual delineations from an independent, young and healthy population, and one manually delineated dataset from STRAT-PARK. It was trained to segment the SN, RN and STN. We used 5-fold cross-validation for training, as the training data was limited. The pipeline is illustrated in Figure 1, and Figure 2 shows an example of the segmentation. The ROIs were coregistered, using Nighres [11], to the space of the other contrast to retrieve the map-values. We used SPM12 to obtain total intracranial volumes (gray matter, white matter and CSF) to normalize volumes of the ROIs [12]. The pipeline is illustrated in Figure 1. From the MP2RAGE and GRE sequences we obtained an R1-map, R2*-map and QSM, utilizing Nighres. For each map we extracted the mean, standard deviation, median, max, min, 5-percentile, skewness and kurtosis for each ROI. For each subject, the statistic values and volumes gave a vector with 225 features. Before inputting the vectors into the classifiers, we scaled each feature using Z-score. We applied various classifiers: support vector machine (SVM), random forest, logistic regression (logreg), XGBoost and K-Nearest-Neighbours (knn) [13, 14]. For each classifier we performed a grid search to optimize hyperparameters.

Even with a variety of classifiers, statistics from R1, R2* and QSM values for the ROIs we selected do not provide enough information to classify subjects as PDP or HC. The accuracies of each classifier are presented in Table 1. For selected ROIs, statistics and contrast, the histograms of mean, 5-percentile, skewness, kurtosis and Jarque-Bera (measure of normality) are presented in Figure 1. Previous results have found a significant increase in the mean QSM value of STN, which is not detected in our cohort. Interestingly, the STN QSM-value in the PD group is similar in the two cohorts, while for the control group it is higher in our study compared to ref [2].

There seems to be bigger intra-class variation than inter-class variation in the two groups. On a group level we know that the putamen volume is significantly lower for PDP than for HC, but this does not show discriminative power. This means that further work should consider other places in the brain than these ROIs to see if more information about the disease can be found. PD is, however, a complex disease and may actually need more complex analysis.

Contrary to previous findings, for our cohort, this combination of statistics, ROIs and contrasts do not contain enough information to predict disease.

Glycogen, a branched glucose polymer, serves as a short-term energy reserve and is primarily stored in the liver and muscles [1]. Disruptions in its metabolism can lead to metabolic pathologies that primarily affect these tissues. In this context, Chemical Exchange Saturation Transfer (CEST) MRI has emerged as a promising non-invasive technique for detecting specific metabolites, including glycogen, in biological tissues [2]. GlycoCEST, which targets hydroxyl group protons in glycogen molecules, has already been applied to liver [3], which has a high glycogen content but also other organs such as brain [4] and skeletal muscles [2]. These studies have primarily focused on in vivo imaging in both healthy and pathological conditions. In this study, we investigate the application of CEST imaging to ex vivo fixed liver tissue at 9.4 T. We begin by optimizing acquisition parameters using glycogen phantoms and subsequently demonstrate the feasibility of glycogen detection in ex vivo rat liver sample and validate it by colorimetric glycogen measurement.

In-vitro experiment: Three tubes containing 0, 50, and 200 mM of bovine glycogen dissolved in phosphate buffered saline (PBS) were prepared for in-vitro validation. To minimize frequency shifts between the phantoms, they were placed inside a larger plastic container filled with Thermasolv CF2 (Fig 1-a). Ex-vivo experiment: Liver tissue (~2.6 g) was harvested from a 250 g, ~12-week-old female Wistar rat immediately after sacrifice. The liver was divided into different portions for complementary analyses. One portion was fixed in formaldehyde to prevent post-mortem physiological changes. During the experiment, the sample was placed in a syringe with Thermasolv (Fig 1-b) and maintained at 37 °C using a temperature-controlled water bath, with temperature monitored via a probe. Colorimetric Quantification of Glycogen: The second portion of the liver was frozen in liquid nitrogen, then crushed prior to extraction. The procedure adapted from Schaubroeck et al. [6], involves alkaline digestion, ethanol precipitation, and resuspension in water. Quantification was performed using a commercial colorimetric assay kit (ab65620; Abcam) based on glycogen hydrolysis and enzymatic detection, with absorbance measured at 570 nm using a microplate reader. MR acquisition & post-processing: Experiments were conducted on a 9.4T/30cm (Bruker Biospin MRI, Ettlingen, Germany) with a cylindrical transmit coil (87-mm intern) coupled with a 4-channel phased-array receive coil. A CEST FLASH sequence with continuous wave saturation was used, with frequency offsets from +3 to –3 ppm (0.1 ppm steps). For ex vivo sample, signal averaging was performed to improve signal-to noise ratio. To optimize acquisition parameters, various B₁ amplitudes and saturation durations were tested, as detailed in Table 1. Initially, the CEST images were reconstructed, corrected for B₀ inhomogeneities using the Water Saturation Shift Referencing (WASSR) method [6] with offsets from +0.6 to –0.6 ppm (0.05 ppm steps) and data were fitted using a two pool Lorentzian fitting approach [7], implemented in MATLAB. After correction, Z-spectra were generated and Magnetization Transfer Ratio asymmetry MTRasym(Δω)= (S(-Δω)- S(Δω))/(S₀) was computed to highlight glycogen presence through its asymmetric CEST effect. MTRasym at 0.7 ppm, corresponding to the glycogen peak [2-4], was then determined using manual ROIs in each phantoms and over the whole visible tissue in ex vivo liver (Fig 1-c,d).

Figure 2 shows the Z-spectra for phantoms and rat liver at saturation time at 1 s and different B1 peak amplitudes. An asymmetry in the Z-spectra, particularly around the glycogen peak in the range of 0.5–1.5 ppm, is observed and increases with glycogen concentration. According to table 2 which summarizes MTRasym values at 0.7 ppm, optimal CEST imaging was obtained with a B1 amplitude of 3 µT and 1.5 s saturation for in vitro experiments, and 2 µT with 1 s saturation for ex vivo sample, both yielding the max glycogen peak with MTRasym of 10.83 % (phantom of 200 mM) and 7.04 % at 0.7 ppm, respectively. Colorimetric quantification revealed a glycogen concentration of 58.95 glucose units per mg of liver tissue, corresponding to 347 mM in glucose units.

CEST imaging has rarely been applied to ex vivo samples [8-9]. Here, we tested and validated the sequence on glycogen phantoms, then tested them on ex vivo rat liver tissue. The observed CEST effect on both experiments matched the glycogen peak reported in in vivo glycoCEST studies (0.5–1.5 ppm) and post-mortem colorimetric analysis confirmed these findings, supporting the feasibility of ex vivo glycoCEST.

These initial experiments serve as a foundational step toward validating the technique before its application to tissues with lower glycogen content, such as skeletal muscle and cardiac tissues. This will support future applications on metabolic alterations in arrhythmias and storage disorders.

Degenerative cervical myelopathy (DCM) is the most common form of non-traumatic spinal cord injury. Cord tissue damage is mainly attributed to cervical compression, while other factors are understudied. Treatment decisions in DCM are challenging due to variable disease progression and a lack of reliable biomarkers[1,2]. Longitudinal structural MRI can provide insight into the temporal pattern of subsequent remote structural tissue changes. This study aims to investigate long-term tissue atrophy using structural T2*-weighted MRI in the cervical cord rostral to the compression site in patients who underwent surgical decompression (operated (OP)) or received conventional non-surgical treatment (non-operated (nOP)).

Data acquisition In this study, 102 DCM patients (mean age±SD: 55.74±11.42 years, 61 males) underwent longitudinal MRI scans of the cervical cord at C3 using 3T MR scanner (MAGNETOM Skyra Fit or Prisma; Siemens Healthineers) equipped with a 20-channel head/neck RF coil. The MRI protocol consisted of sagittal TSE T2- and axial T2*-weighted MRI and clinical examinations at three time points: baseline, 6-, and 12-month follow-ups. Clinical protocol consisted of the International Standards for Neurological Classification of Spinal Cord Injury protocol[3] for evaluation of cervical light-touch and pin-prick scores and total modified Japanese Orthopaedic Association (mJOA) scale[4]. Forty-one DCM patients were operated following baseline clinical assessment. Seventy-two DCM were classified as mild (15≤mJOA≤17) and 30 as moderate DCM (mJOA≤14) at baseline. For comparison, structural MRI data were acquired in 10 healthy controls (HC) (mean age±SD: 45.30±18.54 years, 7 males) at 0, 6, and 12 months. Image processing The cross-sectional area (CSA) of the spinal cord (SCA), grey matter (GMA) and dorsal columns (DCA) were derived from manual segmentations obtained using JIM software (v9). The CSA was averaged across slices covering C3 level. White matter area (WMA) was calculated from subtraction of GMA from SCA[5]. Statistical analysis To ensure greater clinical homogeneity of the compared groups, statistical analysis was conducted in the mild and moderate DCM cohorts separately. For each, two-tailed Fisher’s exact tests were conducted to assess differences in clinical scores at baseline between OP and nOP. An ANOVA was performed to investigate WMA, GMA, and DCA differences between OP, nOP, and HC. Linear models with age as covariate were conducted to determine longitudinal rate of change of clinical scores and tissue-specific CSA, in OP and nOP.

At baseline, WMA, GMA, and DCA derived from T2*-weighted MRI were comparable between OP and nOP groups in both mild and moderate DCM (Fig1). Similarly, clinical scores did not differ between groups in either severity category (Fig2). Both treatment groups (OP and nOP) showed a significant GMA decrease over time, in mild (p = 0.0003) and moderate DCM (p = 0.03) (Fig3). In mild DCM, OP patients showed significant clinical improvement over 12 months post-surgery (Fig4a), while nOP patients showed no significant change. In moderate DCM, both OP and nOP patients improved in mJOA score over 12 months (Fig4b).

Although mild and moderate DCM patients improved in clinical assessment following decompression surgery, MRI revealed DCM-induced long-term grey matter atrophy remote from the compression site in both operated and non-operated patients with mild and moderate DCM. Interestingly, similar patterns of ongoing atrophy have been reported in acute spinal cord injury patients[6,7]. The current findings suggest that, even after successful decompression in DCM, secondary processes may continue to drive atrophy offering insight into DCM progression. Remote grey matter changes rostral the compression site, as observed on structural MRI, may result from anterograde and retrograde degeneration of sensorimotor spinal pathways[8]. The cross-sectional area remains a relatively coarse marker of neurodegeneration and should be interpreted alongside other DCM-specific tools and quantitative MRI sensitive to microstructural changes (e.g., DTI, T1 relaxometry) for a more comprehensive understanding of DCM progression.

There was clinical improvement in patients with mild and moderate DCM following decompressive surgery, yet grey matter atrophy at C3 level was progressive, suggestive of continuing structural degeneration in this cohort. The clinical significance of these findings still needs to be determined, which can be facilitated through longer follow-ups and a comparison to more severely affected individuals. The study’s methodology – combining MRI with clinical assessment across multiple time points – establishes a framework aiming at reaching a more robust understanding of the natural and postoperative course of DCM. Future research will focus on investigating further time points (e.g., 2-year, 5-year follow-ups) to shed light on the long-term evolution of clinical symptoms and atrophy rate.

The potential of amide proton transfer weighted chemical exchange saturation transfer imaging (APTw-CEST) in neuro-oncology has been widely demonstrated [1-3]. Still, clinical implementation remains challenging due to a lack of standardization. Consensus recommendations on data acquisition and analysis for APTw-CEST have been published for brain tumor imaging [4], and reproducibility and repeatability have been proven in various single-site, single-vendor studies [5-8]. Recently, we reported a variability of 12%-57% in APTw-CEST MRI acquired using different scanners and protocols in a phantom [9]. To investigate the ability to distinguish healthy and brain tumour tissue across sites and vendors, we have conducted the first in vivo, multi-site, multi-vendor reproducibility study of APTw-CEST MRI.

Two healthy volunteers were scanned using 3T MRIs at four different sites in the Netherlands that have MRI scanners of three different vendors: Amsterdam UMC (AUMC): Siemens (Vida, Siemens Healthineers, Erlangen, Germany); Erasmus MC Rotterdam (EMC): GE (Premier, General Electric, Chicago, USA); UMC Utrecht (UMCU), Leiden UMC (LUMC): Philips (Ingenia Elition X, Philips Healthcare, Best, The Netherlands). Written informed consent was obtained from both volunteers under local ethical approval. A standard high-resolution T1w image was acquired as part of the scanning protocol at all sites. The acquisition details of the ATPw-CEST sequences are presented in Table 1. The acquisition protocols for AUMC and EMC were chosen based on the consensus recommendations [4]. Acquisitions at LUMC and UMCU were done using the 3D APT product sequence of Philips. Magnetization transfer ratio asymmetry (MTRasym) was calculated using on-scanner processing software for the Philips datasets (LUMC and UMCU) and using in-house processing pipelines in Matlab and the FMRIB Software Library (FSL 5.0, Oxford, UK) for the Siemens (AUMC) [5] and GE (EMC) datasets [6]. Regions of interest (ROIs) were selected using FAST from FSL. The segmented ROIs were white matter (WM) and grey matter (GM). ROI averages and standard deviation (SD) of the MTRasym were calculated per volunteer per ROI. The between-scanner coefficient of variation (CoV) was calculated for each ROI per volunteer by dividing the mean MTRasym of the four scans by the standard deviation of the four scans.

APTw CEST MRI values in both healthy volunteers were low and consistent with expected normative ranges for each acquisition, as seen in the example images in Figure 1. Mean MTRasym signal intensities were between -0.133±1.95 and 0.519±1.71 in GM and between 0.048±0.28 and 0.659±1.36 in WM (Figure 2, Table 2). The between-scanner CoVs range from 40% to 258% across the ROIs and volunteers (Table 2).

The MTRasym values between -0.133% and 0.659% align with reported values for healthy brain tissue [4,5,6]. These values are lower than values for MTRasym in high-grade glioma (HGG), which are reported around 1.6% in the literature [5,6,10]. There is minimal overlap between the SDs of the healthy brain reported here and HGG (from literature), and the range of signal values expected in tumor tissue lies outside the standard deviation of the APTw-CEST values found in three of the four centers in the healthy tissue. The EMC MTRasym values overlap most with the HGG values and differ most from the other three sites. This, in combination with the low intensities measured in this study, has resulted in the high CoVs. Future work will focus on the standardization of post-processing to harmonize MTRasym across the four sites, which is expected to reduce the SD (both within an ROI, as well as between sites), further harmonizing the resulting APTw-CEST quantification. A larger number of volunteers would have improved estimation of the measurement accuracy and variability; however, it would not have allowed us to interpret more from the results, due to the lack of APTw-CEST contrast in the healthy brain. Instead, we have initiated a large, multi-center patient study in which 120 patients with HGG will be recruited across the four sites to give further insight into the in vivo reproducibility of APTw-CEST MRI in the presence of contrast-enhancing pathology.

Combined with the results from the previous phantom study, we hypothesize that, while currently MTRasym differs between sites and scanners, a standardized threshold MTRasym value for diagnosing HGG is likely to be attainable using these acquisition protocols.

Accurate assessment of myocardial viability is essential to reduce the risk of early graft dysfunction following cardiac transplantation. Cardiac magnetic resonance (CMR) is a non-invasive reference technique for characterizing myocardial structure and composition, including edema, fibrosis, and energy metabolism. This study aimed to develop a fast and efficient imaging protocol for tracking myocardial viability over time, potentially enabling surgeons to better evaluate graft quality before transplantation.

Eight hearts were harvested from healthy pigs. Immediately after excision, cardioplegia was administered to arrest cardiac activity and preserve myocardial tissue. Following this, each heart was carefully placed in a container filled with cardioplegic solution to maintain tissue viability. To ensure consistent positioning, the hearts were stabilized inside the container. Imaging was performed using a 1.5T scanner (Siemens Healthineers, Germany), equipped with a 18-channel surface coil. For ¹H-MRS, a PRESS (Point-Resolved Spectroscopy) sequence was used. A single voxel was placed in the interventricular septum. Voxel placement was guided by anatomical localization on four-chamber and short-axis MRI views. Spectroscopic acquisition parameters were: TE=135ms, TR=1500ms, 1024 acquisition points per spectrum, and a spectral bandwidth of 1000Hz. Two types of spectra were acquired: 1) Water-suppressed spectra (32-512 NEX) were used to quantify intramyocardial metabolites, 2) Non-water-suppressed spectra (4-16 NEX) served as references for metabolite quantification. In parallel, T1 mapping of the myocardium was performed using a MOLLI sequence. Imaging parameters were: FOV=150×175mm², a matrix size of 110×128 (determining in-plane resolution), a slice thickness of 8 mm, and TE/TR of 1.21ms and 2.3ms, respectively. Both ¹H-MRS and T1 mapping acquisitions were repeated every hour for a total of 4 hours to assess temporal changes in myocardial viability markers. Spectral data were processed using dedicated post-processing software4. In the resulting spectra, a peak at 1.3ppm5 was interpreted as a composite signal from lactate and lipids, which typically increase in necrotic tissue. Quantification was performed relative to the water signal, and results were expressed as percent concentrations. For T1 analysis, the mid-ventricular short-axis slice was manually contoured once to delineate the endocardial and epicardial borders, and these same contours were propagated across all time points to ensure consistent region-of-interest analysis. The extracted T1 values provided a surrogate marker of tissue integrity and edema, which typically increase with ischemic injury.

Quantification of intramyocardial metabolites using ¹H-MRS was successful in 6 out of the 8 hearts, representing 75% of the dataset. The remaining two hearts were excluded from metabolic analysis due to insufficient SNR in the acquired spectra. An example of a high-quality water-suppressed ¹H-MRS spectrum, acquired one hour after heart explantation, is presented in Figure1. The spectrum clearly shows distinct resonance peaks, particularly around 1.3ppm, corresponding to lactate/lipids. The mean relative concentrations (expressed as percentages of the water signal) of lactate and lipids combined were measured at four different time points post-explantation: Hour 1: 0.024±0.012%, Hour 2: 0.033±0.006%, Hour 3: 0.033±0.010% and Hour 4: 0.037±0.013%. These results show a gradual increase in signal intensity at 1.3ppm over time, which is likely driven by rising lactate concentrations which is a known marker of anaerobic metabolism, suggesting that ischemic processes continue to develop even under cold cardioplegic preservation. In contrast, the lipid component of this peak is expected to remain relatively stable, so the observed increase is interpreted as a surrogate for progressive metabolic distress. In parallel, T1 mapping data were acquired in all hearts. The T1 maps showed spatial homogeneity across the myocardium at each time point, indicating uniform relaxation properties (Figure 2A). The mean native myocardial T1 values increased over time as follows: Hour 1: 727±94ms, Hour 4: 765±89ms. This gradual increase in T1 relaxation time is consistent with the onset of ischemic injury, as T1 values are known to rise in response to edema formation, loss of membrane integrity, and cellular degradation. These findings are illustrated in Figure 2B, which summarizes the temporal progression of T1 values.

This study introduces a combined ¹H-MRS and T1 mapping protocol to monitor myocardial viability in an ex vivo pig heart model. The observed parallel increase in lactate signal and T1 values over time likely reflects ongoing ischemic injury. We hypothesize that beyond a certain threshold—yet to be defined—myocardial viability may become irreversibly compromised. Further studies with larger cohorts are needed before clinical application in human heart graft assessment.

Egg discrimination and rejection is the primary defensive mechanism used by hosts of brood parasites to counteract the negative effects of parasitism, since, in most cases, parasitized hosts would otherwise raise only parasitic offspring, leading to a complete loss of their own reproductive success (1). However, egg rejection occurs at relatively low frequencies. One proposed mechanism underlying discrimination is prolonged learning of the hosts’ own egg appearance, allowing them to recognize and reject foreign eggs (2). This learning process would involve the creation and memorization of a template of their own eggs, which could be used to identify mismatched eggs across multiple breeding attempts. It has been demonstrated that the hippocampus is the brain region responsible for memory, while the nidopallium is implicated in learning processes (3). The aim of this study was, for the first time, to compare these two brain regions between egg-rejecting and egg-accepting females using in vivo Diffusion Tensor Imaging (DTI) and assess anatomical brain volume changes between sexes.

Wild magpies were trapped in the town of Calahorra, Granada, Spain. Then they were moved to the preclinical MRI facility of the University, where they were anesthetized with 2% isoflurane in oxygen via a veterinary mask for induction and maintained under 1–2% isoflurane using a nose cone during scanning. Animals were positioned supine, restrained with tape, and placed over a water-heated blanket. MRI scans were acquired on a 7T Bruker Biospec scanner (20 cm bore, Paravision 360 v3.5) using a 1H rat brain surface coil and an actively detuned transmit-only resonator. Anatomical images were obtained with a variable flip angle 3D-T2-weighted RARE for brain volume assessment. For diffusion MRI, field mapping were performed prior to acquisition. DTI data were acquired using a two-shot EPI sequence with 60 diffusion directions (b=1000 s/mm²) and 10 b0 images. (TR/TE=7000/25 ms, δ/Δ=4/12 ms), total study time ~75 min (figure 1). 3D-T2WI were denoised employing 3D adaptative multiresolution nonlocal means denoising algorithm (4) and DWIs were processed using mrtrix3 software (5). Regions of interest involved in learning, problem solving, and memory (hippocampus, mesopallium, nidopallium, and tectum opticum) were delineated blindly and manually by a an avian neurobiologist, and mean value of the ROI was employed in statistical t-test.

Preliminary data point that females (n=6) exhibited larger brains than males, even after normalizing for body weight (figure 2), suggesting potential sex-based differences in neuroanatomical organization (6). Preliminary significant differences in DTI-derived parameters (including fractional anisotropy (FA), mean diffusivity (MD), and axial diffusivity (AD)) were found between females that rejected parasitic eggs and those that did not (n=4 per group). Specifically, higher FA, MD, and AD values were observed in non-rejecting females, which may reflect differences in hippocampal microstructure (figure 3).

These findings, together with the hippocampal differences observed between groups (and the absence of differences in the nidopallium), suggest that the ability to discriminate between own and parasitic eggs could be linked to local neuro-architectural variations. One plausible explanation is a greater organization of interneurons in the hippocampus of rejecting females, supporting enhanced processing capabilities related to egg recognition. Previous studies have shown that DTI metrics can reflect underlying neuronal composition in avian brain (6,7), supporting the hypothesis that rejecting females may exhibit a more refined brain microstructure.

Although further behavioral recorded video-data are needed to be analyzed to confirm this structure-function relationship, our findings point toward a neurobiological basis for individual variation in parasitic egg discrimination. We successfully established an in vivo MRI and DTI protocol for wild magpies, enabling the investigation of neuroanatomical correlations of egg rejection behavior. Our results reveal significant microstructural differences in the hippocampus between egg-rejecting and non-rejecting females, suggesting a potential neurobiological basis for individual variation in parasite egg discrimination.

Diffusion is a key technique in Magnetic Resonance Imaging for probing the brain’s microstructural organization. It provides critical biomarkers such as fractional anisotropy (FA), or tractography, which enables the reconstruction of structural connectivity between brain regions. These analyses typically rely on Diffusion Tensor Imaging (DTI) sequences. To achieve high-resolution 3D maps in small animals and employ advanced models such as High Angular Resolution Diffusion Imaging, a large number of diffusion gradient directions must be acquired. This requirement leads to prolonged acquisition times—often exceeding 10 hours—thus precluding in-vivo applications. To address this limitation, one strategy is to reduce the number of acquired measurements and apply Compressed Sensing reconstruction techniques to recover relevant diffusion contrasts. Among these, low-rank reconstruction methods exploit the redundancy across neighboring voxels by assuming similar signal behavior within local spatial patches, without requiring explicit signal models. Alternatively, subspace reconstruction methods incorporate prior knowledge of signal behavior through simulations, enabling data recovery by projecting onto a learned signal subspace. This study proposes a proof of concept demonstrating that the combination of an advanced reconstruction framework with optimized and original sampling schemes enable substantial acceleration of 3D diffusion MRI acquisitions of the mouse brain while preserving image quality and the reliability of derived diffusion metrics.

Materials: Experiments were performed on a 7T Bruker BioSpec system (Ettlingen, Germany) equipped with a volume resonator for excitation, and a 4-element (2×2) phased surface cryoprobe for signal reception. Sequence: A 3D Spin-Echo sequence was applied with a 156µm isotropic resolution, matrix: 128x128x96, 30 directions, 5 b0, total acquisition time=12h. Ex vivo healthy-mouse brains were used, after brain extraction and PFA fixation. Acceleration: Acceleration was performed by a posteriori removing data based on a Variable-density Poisson Law along the phase and partition axes. The center of the 3D k-space was left fully encoded. A random sampling was performed along the diffusion directions. Different acceleration factors (from 6 to 12) were tested. Reconstruction: For reconstruction we used two methods: Low Rank reconstruction and Subspace reconstruction: - For Low Rank reconstruction we used the Local Low Rank (LLR) with a wavelet regularization. We optimise both reguliization parameters for low rank and wavelet - For the subspace reconstruction, two dictionnaries were built. One based on simulation based on the DTI model by creating a large number of ellipsoids with different eigen values and different orientations. The other is based on the fully sampled central data, by reconstruct low resolution images. The simulation and reconstruction have been perfomed using Julia and the package MRIReco.jl and the Bart ToolBox (Berkeley Advanced Reconstruction Tool) Image analysis: FA maps were reconstructed using Dipy on Python to validate the acceleration process by comparing the standard deviation value between undersampled and fully sampled data as a reference.

For the low-rank reconstruction, an ℓ1-regularization was applied in the wavelet domain along the three spatial dimensions, while a Locally Low-Rank regularization was performed along the diffusion dimension. The regularization parameters were optimized by minimizing the Mean Squared Error between the accelerated acquisition and a fully sampled reference. (Fig.1) For the subspace reconstruction : the optimal number of subspace dimensions was determined by testing various values on fully sampled data. The results indicated that, although the singular values decay rapidly, a relatively large number of components is required to ensure accurate reconstruction (>10). (Fig.2) Comparison of the reconstructions from under-sampled data using both the subspace and low-rank approaches suggests that, under the current simulation setup, both methods yield comparable results even if subspace reconstruction tend to offer a more accurate reconstruction. Less than ± 1% error has been calculated beetwen the mean FA value of the fully sampled corpusm callux (0.488 ± 0.097) and an acceleration factor 6 subspace reconstruction with a simulated dictionnary (0.489 ± 0.095). The Low Rank reconstruction with the same acceleration factor shows a difference of 5%.

An acceleration factor of 6 can be applied without compromising image quality for both reconstruction approaches (Fig.3) corresponding to an acquisition time of approximately 2 hours. FA maps confirmed that accelerating the acquisition by a factor of 6 did not alter the resulting diffusion metrics. However, the current acquisition time still remains relatively long, and further optimization is necessary to validate the feasibility of using higher acceleration factors.

Real-time monitoring of microwave liver ablation (MWA) using MRI thermometry can be hindered by susceptibility artefacts [1,2] caused by boiling [3]. The latter provoke unphysical temperature deviations that prevent lesion size prediction [4,5,6]. This study proposes a correction based on removing the contribution of the susceptibility artifact using subvoxel sources of susceptibility.

Experiments: The protocol was approved by the ethics committee for animal experimentation CEEA50 (France). MWA (N=23) were performed using an AveCure microwave system (14-gauge large antenna, MedWave, USA) on seven pig livers. Energy delivery was set to 7’30 with a target temperature of 80°C. Acquisitions and dosimetry: Acquisitions were performed under respiratory gating during the most stable part of exhalation using multi-shot gradient-echo echo planar imaging sequences on 1.5T and 3T Siemens. The voxel size was 2.3 x 2.3 x 3.0 mm3, slice gap 1.5 mm with 7 slices. At the end of the experiment, gadoteric acid (0.5 mmol/kg, GE Healthcare) was injected intravenously, and 3D T1-weighted images were acquired to visualize the non-perfused volumes at a resolution of 1.2x1.2x1.2 mm3. Temperature calculation was performed using the PRFS method, followed by a spatial-temporal drift correction. The proposed correction of susceptibility artifacts was then applied. The cumulative TD based on the Sapareto equation was computed, taking an equivalent dose of 240 min at 43°C (CEM43) as the lesion volume (‘TD ablation zone’) Susceptibility correction methodology: Figure 1 shows a schematic overview of the method. It consists in generating a susceptibility distribution that reproduces the main stable component of the artifact through time. The magnetic susceptibility artifact simulation used the dipole approximation of the field perturbation caused by local changes of magnetic susceptibility. The susceptibility values were determined at a finer resolution (subvoxel) and using adjacent slices contribution by solving a linear inverse problem which was set to nullify the artifact. Temperature data and lesion size analysis: The data analysis was performed retrospectively, aimed at assessing the ability of the method to reproduce the artifacts and correct the temperature deviations. Second, to demonstrate the advantages of this method under actual treatment conditions, comparisons were performed with the postprocedural ‘T1w ablation zone’ against two ‘TD ablation zones’: one given by the uncorrected data and one given by the proposed correction.

Out of seven pigs with 23 MWA, 3 (13%) MR temperature acquisitions were excluded due to low signal-to-noise ratio (SNR). Susceptiblity artifacts caused by boiling were observed in 11 (48%) acquisitions. Figure 2 presents a selection of corrections over six MWA with varying degrees of: artifact intensity, shapes of the artifact, orientations of the ablation probe in respect to the slice and B0 magnitudes (1.5T and 3T). Figure 3 compares the time evolution of temperature and thermal dose for a slice adjacent to the MW probe that was impacted by a positive lobe of the artifact. The 10°C offsets visible in the marked voxels are removed after the correction, causing a considerable impact on the dosimetry estimate, as the TD threshold is no longer reached in those voxels. Figure 4 displays the T1w ablation zone segmented from post-contrast images versus TD ablation zones with and without the correction. The first case shows a discontinuous lesion shape due to the presence of a vessel, while the other shows an ellipsoid-shaped lesion. Without correction, an overestimation of lesion size up to 9 mm (two slices of 4.5 mm thickness) in a minor axis was observed. The geometric and overlap metrics indicate an increase of agreement between corrected-and-T1w measures versus uncorrected-and-T1w: (i) overestimation of lesion volume (T1w-TD) was reduced in average from -1.02 to -0.59 cm³ (42% reduction), most noticeable in Y (26% reduction in average from -3.46 to -2.55 mm) and Z axis (17% reduction in average from -3.20 to -2.67 mm); (ii) median volumetric Dice increased by 4.8% (0.62 to 0.65).

The use of subvoxels was motivated to diminish the main pitfall in the dipole model: for positions close to the point of non-zero susceptibility, the kernel will induce deviations that grow without bounds. The prior selection is a potential limitation. A perfect correction would require a time dependent simulation. However, the artifact is relatively stable through time.

Comparison with T1w imaging showed improvements in prediction of lesion volume. The proposed algorithm takes into account partial volume effects and the contribution of adjacent slices. It possesses the necessary adaptability for simulating a diverse range of deformed dipole-like artifacts seen in experimental data. It could also be applied to radiofrequency or laser energy.

The hMRI toolbox is an SPM-based, collaboratively developed open source software on GitHub for generating quantitative MRI (qMRI) maps sensitive to myelin and iron content [1]. It produces high-quality multiple quantitative parameter maps (R1, R2*, PD, and MTsat) that facilitate advanced analyses, precise subcortical delineation, and provide critical inputs for in vivo histology via biophysical models [2]. Here we highlight the upcoming 1.0 release, containing updates to tissue weighted smoothing and mask creation as well as new capabilities such as advanced image denoising, the quantification of errors on the qMRI estimates, and a correction for motion-related image degradation in statistical analyses (QUIQI) [3-6]. These features seamlessly integrate with the existing DICOM import, map creation and qMRI map processing (spatial processing) modules (Fig.1). Furthermore, the toolbox can now run as a standalone application, is available in Neurodesk [7] and is supported by the upload of the public demo dataset [8,9] to GitHub by utilizing the Git “large file storage” (LFS) (Fig.2), promoting reproducible research and community engagement.

The updated hMRI toolbox has been re-engineered as a standalone application that can run on platforms such as Neurodesk, allowing users to operate the toolbox without relying on a MATLAB license. At the start of the processing workflow (Fig.1) the DICOM Import module facilitates the conversion of raw DICOM files to NIfTI format. Following this, a state-of-the-art denoising module [3]—integrated via a custom Java-MATLAB interface— provides access to a set of advanced noise reduction techniques [10-12] (Fig.3). Subsequently, the Create Maps module converts the preprocessed data into quantitative maps. The Error Maps module (Fig.4a), quantifies voxelwise errors on the qMRI estimates and can adaptively combine multiple sets of qMRI data to downweight regions with high error. In group-level analyses, the QUIQI module (Fig.4b) accounts for the level of motion-related degradation of each individual map to ensure the validity of the statistical results. The index of motion degradation estimated by the Create Maps module of the toolbox [13] or another index of the user’s choice can be used. Enhanced metadata support is provided through automatically generated JSON sidecar files, and the inclusion of a public demo dataset [8] (Fig.2) allows users to validate and refine the workflow.

The processing workflow in hMRI toolbox 1.0 provides substantial improvements in both usability and data quality control. The standalone version successfully broadens accessibility by eliminating the dependency on MATLAB licensing, simplifying deployment across varied computational environments. Users can now employ a seamless framework beginning with the DICOM Import module, followed by advanced denoising, and subsequently the creation of quantitative maps via the Create Maps module (Fig.1). This integrated approach ensures that the output maps exhibit markedly reduced noise [3] (Fig.3). In addition, the Error Maps module provides voxelwise assessments of map quality that allow the adaptive combination of maps acquired across different repetitions. The QUIQI module allows the mitigation of motion-related image degradation to ensure the validity of statistical results in group-level analyses (Fig.4). Testing on the public demo dataset [8] confirmed that these new modules can enhance users’ capability to conduct analyses of microstructural brain tissue change from qMRI data.

The recent enhancements in the hMRI toolbox epitomize a strategic refinement in the processing of quantitative MRI data. The addition of state-of-the-art denoising techniques, coupled with the detailed Error Maps module and the QUIQI module for motion compensation, further refine the map creation process. The ability to use these modules in a workflow—beginning with DICOM Import, progressing through denoising and map creation, with the option to apply error assessment, motion mitigation, and dedicated spatial processing—creates a cohesive processing tool that significantly enhances data quality and reproducibility. The standalone application expands accessibility to diverse computational environments like Neurodesk.

hMRI toolbox 1.0 is a significant advancement in quantitative MRI analysis. By combining reliable DICOM import, map creation and spatial processing (Fig.1) with state-of-the-art denoising [3], meticulous error mapping (Fig.4a), and effective motion mitigation (Fig.4b), the toolbox delivers robust qMRI biomarkers critical for in vivo histology. Its utility spans a diverse range of applications—including developmental and aging studies [14,15] and explorations of neurodegenerative or neuroinflammatory conditions [16,17]—cementing its role as a powerful resource for enhancing research efficiency and fostering open source software development and reproducible neuroimaging methodologies.

Due to the phylogenetic proximity and the strong homology between the human and non-human primates’ brain, investigating the neuroanatomy of the macaque brain helps further our understanding of the human brain and the development of novel therapeutic approaches. Diffusion-weighted MRI is a powerful non-invasive tool to investigate the microstructure of the brain[1]; however, it remains challenging to acquire in vivo data in macaque monkeys. Their brain size requires higher spatial resolution, but their mensuration generally prevents the use of preclinical scanners designed with higher magnetic fields. Here, we propose a replicable method to acquire sub-millimeter isotropic diffusion MR imaging on anesthetized primates with a standard clinical MRI scanner. We also make available to the scientific community a dataset on 9 macaque brains comprising raw diffusion and anatomical MRI data, the associated templates, and the data curated for MR tractography.

We acquired data from 5 male and 4 female healthy macaque monkeys (Macaca fascicularis) aged between 3 and 23 years old at the time of experiments. All experiments were performed with a 3T Prisma Siemens MR scanner, using the vendor’s body coil for transmission and a 64-channel head coil for signal reception, as commonly used for human acquisitions. The animals were positioned in supine position, with their head maintained straight in the MRI coil using a shape memory head support (MOLDCARE head cushion). Each animal underwent two MRI sessions dedicated to diffusion imaging (2h36min) and anatomical imaging (1h11min). The animals were intubated and maintained under anesthesia using 1-3% isoflurane mixed with pure oxygen. Diffusion-weighted acquisitions were obtained with an in-house MR pulse sequence, a 3D spin-echo segmented EPI, designed to maximize SNR and provide high isotropic resolution while maintaining the total acquisition time below 3 hours for the animal's wellness. Images were reconstructed using the standard vendor’s reconstruction. Key parameters were: Axial orientation, TR/TE= 700/64ms, receiver bandwidth= 776 Hz/pixel, Partial Fourier 7/8, 4 segments, FOV=100 x 102.9 x 61.6 mm3, nominal isotropic resolution of 0.7 mm3, 32 diffusion encoding directions at b=1000s/mm2, two acquisitions with b=0s/mm2 (named b0) and opposite phase encoding directions. In addition, high-resolution T1-weighted (MPRAGE) and T2-weighted (SPACE) acquisitions with isotropic resolution of 0.5mm3 were collected in a separate session with Siemens product sequences. All images were processed using versaFlow[2], an openly available pipeline designed to preprocess and analyze diffusion MRI data from macaque brains. Diffusion MRI preprocessing included filtering for background noise, magnetic susceptibility, eddy currents, motion, and intensity bias. Diffusion profiles were obtained by Diffusion Tensor Imaging (DTI)[3] and Constrained Spherical Deconvolution (CSD)[4]. Probabilistic local tractography[5], [6] was performed with a GPU implementation in the Sherbrooke Connectivity Imaging Lab (SCIL) Python dMRI processing toolbox (Scilpy) on the whole brain. The common template space was generated from all 9 subjects using a multi-modal registration approach available in ANTs. The parcellation of the INIA19 atlas[7] was used as regions of interest (ROIs) to extract some bundles of interest, namely the corpus callosum (CC), the anterior commissure (AC), and the pyramidal tract (PyT). Quality control assessment was performed for the 9 datasets at each step of the versaFlow pipeline, using the Diffusion MRI Quality Check in the Python (dmriQCpy) toolbox[8].

T1, T2 and diffusion weighted images acquired for each of the 9 animals are shown in Figure 1. The CNR of the T1 weighted image and the SNR of b0 were equivalent for all animals (mean/std=0.7/0.1 and 4.1/0.3 respectively). No correlation was found between age/weight and quality metrics (R<0.55), suggesting that the image quality is not driven by the subject's age or weight in this dataset. DTI metrics are available for all subjects (Fig.2) along with the averaged template of the weighted images and the computed metrics (Fig.3).

We could observe high variance of the data (Fig.3) in the basal ganglia, which can be explained by iron accumulation across age[9]. This variance was mainly observed in the T2 weighted and b0 images, minimally affecting the template of DTI metrics. The extracted bundles followed the known macaque monkey neuroanatomy (Fig.4): The color-coded concatenation of these callosal streamlines, displayed in the bottom row, corresponds to the frontal, parietal, occipital, and temporal topological organization of the homotopic callosal fibers as previously described[10].

Interested users will be able to replicate the acquisitions on other macaque monkeys and further enrich the MADI@700 dataset. Moreover, the current dataset, already processed and curated for tractography, may be used to investigate microstructural biomarkers.

The pituitary gland (PG) is a critical endocrine component. It comprises the anterior and posterior lobes (with the less distinct pars intermedia) that serve distinct functions. The anterior PG regulates several neuroendocrine axes, including the hypothalamic-pituitary-gonadal, hypothalamic-pituitary-adrenal, among others. The posterior PG is dedicated to storing and releasing hormones. The PG's size and structure are subject to various influences such as puberty, nutritional status, medication use, and psychiatric disorders, highlighting the importance of detailed volumetric studies. While most studies focus on the entire PG volume, the functional and structural differences between the anterior and posterior lobes suggest that separate analyses could yield more precise insights. Quantitative analysis of brain MRI relies on accurate anatomical segmentation, traditionally achieved through labour-intensive manual labelling. Automated methods offer a more efficient alternative. However, the absence of a standardized PG atlas impedes research in this area. This study aimed to develop optimized spatial normalization protocols for the anterior and posterior PG in healthy young adult females, establishing a crucial reference for future research, particularly in psychiatric conditions prevalent in women, such as anorexia nervosa.

26 healthy young female volunteers (mean age 23±3 years) with normal body mass index (18.5-25.0 kg/m2) were included. Participants had no history of brain or psychiatric disorders and were not using any medications or drugs. High-resolution T1-weighted MRI scans were acquired using a 1.5 T Siemens Sonata system. Manual delineation of the anterior and posterior PG lobes was performed using FSLeyes software. Given the challenges in normalizing the PG within a standard stereotactic space, a three-step optimized spatial normalization procedure was employed. The ICBM152 template in SPM12 was used in the first step, involving linear and non-linear transformations with specific parameters (7×8×7 discrete cosine transform basis functions and 12 iterations). Deformation fields were applied to normalize the binary pituitary delineations, resulting in an initial non-optimized pituitary atlas. A second normalization step, limited to the PG mask, was performed to refine the alignment, addressing the limitations of standard brain masks that exclude the PG. This approach, masking the cost function, optimized the normalization specifically for the PG. A third normalization step further improved the atlas. Atlases for the anterior and posterior PG were created using the spatial transformations from the third normalization step.

One participant presented with an intrasellar arachnoidocele, a common condition, and was included in the analysis to reflect population variance. The mean total pituitary volume was 705±88 mm3, with anterior and posterior volumes of 614±82 mm3 and 91±20 mm3, respectively. The probabilistic atlases showed smooth contours and clear boundaries between the anterior and posterior lobes. The global pituitary atlas showed an 80% overlap with manual delineations for the DICE index and 67% for the Jaccard index. For the anterior pituitary atlas, these values were 77% and 64%, and for the posterior PG atlas, 62% and 47%, respectively. Manual delineations closely aligned with the atlas-based estimation of 688±78 mm3, with a non-significant mean relative volume difference of 2.4%. A significant correlation (R=0.44, P=0.025) and good agreement were observed between the two methods. Anterior PG volumes were also similar between manual (615±82 mm3) and atlas-based (607±73 mm3) measurements, with a 1.3% mean relative volume difference and a significant correlation (R=0.43, P=0.027). Although posterior PG volumes were similar between methods (manual: 91±20 mm3, atlas: 92±10 mm3, 1.1% relative difference), the correlation was not significant (R=0.09). There were no significant differences in average anterior-posterior (AP) ratios between manual and atlas-based methods (5.6% mean relative difference), but a strong correlation was observed (ρ=0.54, P=0.004).

This study successfully developed the first maximum probability atlas of the pituitary gland, specifically tailored to young adult females. This atlas is a valuable tool for research, particularly in psychiatric disorders with a high prevalence in women, such as anorexia nervosa. The ability to differentiate between the anterior and posterior PG enhances the precision and robustness of research outcomes. The measured PG volumes are consistent with the range reported in other studies, with the reduced variance potentially attributed to methodological improvements or the focus on a homogeneous young female group.

The atlas demonstrates good agreement with manual delineations and shows substantial concordance for the entire PG and anterior lobe volumes, as well as AP ratios.

Superparamagnetic iron oxide nanoparticles (IONs) have excellent magnetic properties for applications as MRI contrast agents (CA). Relaxometry can improve quantification of ION concentration or spatial distribution, relevant for biomedical research or medical diagnostics. However, the exact relation between changes in relaxation time and CA properties remains elusive [1]. Here, an open-source simulation tool was implemented that enables characterization of relaxation time changes for ION-based CA in MRI.

A Monte Carlo simulation tool was implemented for simulations of the changes in transverse relaxation times T2 and T2* caused by IONs with given properties (Fig. 1) [2]. Inside a cubic voxel (10^-6 – 10^-3 mm^3) IONs were modeled as randomly distributed impermeable spheres, with corresponding dipole fields. Concentration, magnetization, size, and coating of IONs can be defined manually. Diffusion of 10,000 water protons was modeled as random walk for 100 ms. The phase of each water proton spin was accrued over a given number of diffusion steps (10^5 – 10^7) through the local dipole fields. The MR signal at each diffusion step was obtained with an exponential fit of the signal. For T2 simulations, a multi spin echo sequence was implemented, with phase inversions of all spins after each refocusing pulse. The simulation tool was tested on two ION-based CA, Ferucarbotran (Bayer AG, Leverkusen, Germany) and nanohybrid magnetic nanocubes (MNC@OH) [3]. Ferucarbotran consist of magnetite crystal clusters inside a coating of carboxydextran, with a median hydrodynamic radius of 22.5 nm and magnetization of 380 kA/m (Fig. 2) [4, 5]. MNC@OH were prepared with a ligand exchange procedure using tetramethyl ammonium hydroxide, which results in a median hydrodynamic radius of 177.2 nm and magnetization of 385 kA/m. For measurements of T2 and T2*, CAs were prepared in agarose phantoms with iron concentrations of 90, 180, 400 µmol/L and 140, 300, 1000 µmol/L for Ferucarbotran and MNC@OH, respectively. Measurements were performed on a 9.4 T animal scanner (Bruker BioSpec 94/20). For T2, a multi spin echo sequence (TE,min = 6.5 ms, echo spacing = 6.5 ms, 16-25 echoes, TR = 2000 ms) and for T2* an ultra short echo sequence (TE,min = 0.45 ms, TE,max = 40 ms, TR = 120 ms, flip angle = 20°) was used. Simulations of the CA have been performed for three iron concentrations corresponding to the measurements and two approaches for modeling the CA size were tested (Fig. 2C). In the first approach, IONs with uniform radii were simulated, using the median hydrodynamic radius. In a second approach, the distribution of hydrodynamic radii for MNC@OH, and of the magnetite clusters with a coating layer for Ferucarbotran were considered for the simulations. To account for potential particle aggregation of MNC@OH, simulations were also performed with various larger hydrodynamic radii.

Measurements and simulations reproduced the linear correlation between concentration and relaxation rates R2 and R2* (Fig. 3). For Ferucarbotran, the simulated R2 and R2*, when using the uniform median radius, were twice as high as the measured values. When the radius distribution and coating were considered, simulation results were in good agreement with the measurements. For MNC@OH, the simulated relaxation rates were similar, irrespective of the approach for modeling the CA size. Deviations between simulated and measured relaxation rates were lowest when the hydrodynamic radius distribution was considered. Simulated R2* fit well to measured values, but for R2 substantial differences were observed. Cryo-electron microscopy (EM) revealed substantial aggregation of MNC@OH (Fig. 4A). Simulations accounting for these larger hydrodynamic radii showed that R2* was independent of the radius, whereas R2 decreased with increasing radius (Fig. 4B). Simulations with Rh = 400 nm reproduced the measured R2* and R2 for MNC@OH (Fig. 4C).

The developed simulation tool, with an appropriate approach for modeling CA particle sizes, was able to reproduce experimentally measured R2 and R2* for different ION-based CA. For comparisons of simulated and measured relaxation times, the radius distributions of the CA have to be considered in the simulations. Differences between simulations and measurements of the MNC@OH particles were related to aggregation, confirmed by cryo-EM. The median hydrodynamic radius of these aggregates could be estimated from simulations alone.

In this study, an open-source Monte Carlo simulation tool was implemented and validated for simulations of ION-based CA in MRI. This tool can be used to characterize the influence of changes in specific parameters of CA on relaxation time, which was confirmed here by revealing particle aggregation of MNC@OH. Comparison of measured and simulated relaxation times of CA with distinct properties may enable exact quantification of CA concentration and spatial distribution in complex environments, such as tissues of interest.

Proton magnetic resonance spectroscopy (¹H-MRS) offers non-invasive imaging of brain metabolism [1]. In preclinical neuroscience, rodent models are widely used for understanding disease mechanisms and progression, treatment planning, and studying aging-related diseases [2,3]. Although advanced tools for MR spectroscopy analysis and visualization have been developed for clinical MRS data [4], their applicability to preclinical research remains limited due to differences in data format, anatomical scale, and acquisition protocols. ORYX-MRSI, originally developed for clinical ¹H-MRSI data, integrates features such as spectral visualization, co-registration, and atlas-guided statistical evaluation [5], but requires adaptation for use in rodent studies. The proposed toolbox, Preclinical Oryx-MRS, offers a user-friendly graphical interface and supports key functions such as spectral loading and visualization, co-registration, image registration, segmentation, atlas-based quantification, and metabolite concentration visualization, specifically tailored for rat brain ¹H-MRS data.

Preclinical ORYX-MRS was programmed in MATLAB 2024a (The MathWorks Inc., Natick, MA). The software currently accepts various user-provided input files, including a raw ¹H-MRS data file in .mrd format, an LCModel output .coord file, and a reference anatomical MRI. The toolbox currently includes five main modules: Load Data, Co-registration, Registration, Segmentation, and CRLB & FWHM & SNR. Load Data: This module reads and visualizes both raw ¹H-MRS .mrd files and LCModel .coord files. Users can toggle between raw and processed spectral visualizations. Co-registration: This module creates a volume of interest (VOI) mask using offset, angle, and size information from the .mrd file. The generated mask is then co-registered to the anatomical reference MRI in NIfTI format. The final mask file is saved as a NIfTI image, and SPM [6] is used to visualize the reference image overlaid with the VOI mask. Registration: This module registers the anatomical reference MRI to the SIGMA rat brain atlas [7] using FMRIB's Linear Image Registration Tool (FSL FLIRT) [8]. The resulting transformation matrix is then applied to the VOI mask. Both the registered anatomical image and the mask are then saved in NIfTI format. Segmentation: This module uses the anatomical reference image and the SIGMA rat brain atlas segmentation masks to segment white matter (WM), grey matter (GM), and cerebrospinal fluid (CSF) using the SPM toolbox. It then calculates the regional volume fractions and interactively visualizes the results. The segmentation module adapts modified functions from Osprey [9], which was originally developed for clinical MRS data. CRLB & FWHM & SNR: This module extracts FWHM and SNR values from the LCModel .table file and plots the CRLB and concentration values for selected metabolites. Toolbox Assessment: A rat brain was scanned on a 7T preclinical MR scanner (MR Solutions, Guildford, UK). T2-weighted anatomical MRI (TR/TE = 5000/45 ms, slice thickness = 1 mm) and 1H-MRS data (PRESS (Point Resolved Spectroscopy), 3x3x4 mm3 voxel, TR/TE= 3000/24 ms, 2048 time points, 256 averages) were acquired. The .mrd formatted MRS file, .nifti formatted T2w MRI, and a .coord file from an open source [10], were used to evaluate the toolbox.

Figure 1 shows the interface of the toolbox, highlighting five of its modules and shows visualizations of a raw ¹H-MRS .mrd file and an LCModel .coord file. Figure 2 displays the VOI mask co-registered to the T2-weighted rat brain MRI, which was then registered to the SIGMA atlas. Figure 3 illustrates the segmented WM, GM, and CSF, their corresponding volume fractions, and the FWHM & SNR plots.

Preclinical ORYX-MRS is a user-friendly ¹H-MRS data analysis toolbox developed as an extension to the ORYX-MRSI platform. It currently features five modules for reading and visualizing MR spectral data, creating and co-registering VOI masks, registering anatomical MRI and masks to the SIGMA atlas, segmenting WM, GM, and CSF, calculating tissue fractions, and plotting LCModel-derived quality metrics. Future updates will include LCModel-based quantification of .mrd files and a region-based data analysis module, and data distribution plots.

Preclinical ORYX-MRS provides a solution for the analysis and visualization of preclinical ¹H-MRS data in rat brain studies. By integrating essential functions such as spectral visualization, VOI mask generation and registration, tissue segmentation, and LCModel-based quality metric extraction, the toolbox will aid preclinical MRS workflows.

Ischemic stroke (IS) is a major cause of disability and death worldwide. Endovascular Thrombectomy(EVT) has proven to be highly effective, and is now a standard clinical intervention [1]. Baseline(D0) and 24hrs follow-up(D1) MR or CT data are routinely acquired. MR angiography (MRA) is often included in the routine protocol, and sometimes with MR perfusion to confirm diagnosis or treatment decision [2]. A safe and efficient EVT requires a strict management timeline and highly skilled interventionalists. Training is a key element of performance and process improvement due to the need of complex microcathether navigation into the intracranial circulation. Recent international recommendations(WFITN) integrated simulation in the training procedure [3]. The planning phase of EVT may also benefit from simulation with predictive capabilities, e.g. to plan catheter navigation (mechanical simulation of the interventional devices), to improve perfusion imaging or to predict reperfusion (computation of fluid dynamics). New brain vasculature segmentation and modelling methods are therefore necessary to provide geometric boundary conditions for numerical predictive simulation, and the dedicated real-world patient image database is necessary. While several brain vasculature databases exist [4], real-world brain MRA database from acute ischemic stroke phase is missing. Clinical images are often difficult to obtain. In this work, we present how we constructed such a MR angiography database, named PreSPIN (Predictive Simulation for the Planning of Interventional Neuroradiology procedures), from real-world clinical images. Vessel segmentation obtained from fine-tuned deep learning segmentation algorithm is also included.

Database specifications: The database construction is based on the following inclusion criteria to balance between the clinical scenario and algorithm development requirement: 1).MR field strength used in everyday practice; 2).MR 3D Time of Flight(TOF) images with diagnostic quality; 3).Patients with both D0 and D1 TOF availability; 4).D0 MR perfusion data availability. Thanks to our research image data platform ArchiMed[5], EVT related multi-centric clinical trials are archived together including metadata, and candidate patients for this database can be easily identified. Quality control was conducted and saved as metadata. Database construction procedure: Selected DICOM headers (magnetic field strength, image resolution, and brain coverage) of all IS patient MR TOF data were analysed to characterise in the year (2021) before building the database (2022).Only exams with a diagnostic image quality noted at least as “good” were selected. They were then sorted according to: D0 and D1 availability, then D0 availability; decreasing voxel isotropic ratio (VIO, defined in Figure 1 c). Exams with insufficient 3D TOF coverage were also excluded. This was assessed using MIP visualization in 3D Slicer, with a preference for full head coverage, and at minimum full visibility of the Circle of Willis and major arteries of the anterior and posterior circulation. 3D TOF data from EVT patients did not cover the full head and was considered non-identifiable. The data were anonymised and saved in BIDS format. They were also detached from clinical trials. A pretrained deep learning (DL) algorithm [6, 7] was further fine-tuned and applied to this database for future expert modification.

A total of 210 MR 3D TOF data were identified (acquired between 09/2021 and 03/2023), involving 161 patients (mean age: 65yrs; 104 male, 55 female), with a subset of 49 patients (mean age 63yrs, 32 males vs 17 females) who had both D0 and D1 data. Data acquisition was performed on 1.5T and 3T systems in a ratio of 65% to 35%, consistent with our investigation for data in 2021(68% 1.5T, 32% 3T). Additionally, 54 D0 MR perfusion datasets were identified, among which 26 patients had corresponding D0 data included. Initial segmentation of all TOF data were incorporated into the database. Database characteristic histograms are shown in Figure with the age, image resolution and VIR distribution.

Real-world patient image data are important for algorithm development. We have shown that our database reflected clinical practice. Future updates will include expert corrections to the initial DL algorithm output segmentation.Indeed, any segmentation is error prone, and what is considered a good segmentation might depend on the targeted clinical application [8]. The PreSPIN database has the potential to serve as a collaborative space where multiple experts can provide their correction to the same exam, fostering future algorithm development.

The PreSPIN MRA database is well-structured and close to real-world clinical imaging data. It can serve as a benchmark for testing different brain vascular segmentation and modelling algorithms.

Alzheimer’s disease (AD) and frontotemporal dementia (FTD) are complex and heterogeneous neurodegenerative disorders, each characterized by variability in clinical presentation, genetic factors, and neuroimaging profiles [1], [2], [3], [4]. The overlap in symptoms and imaging features between AD and FTD complicates the diagnosis [5]. To address this challenge, we propose PatientSpace, a graph-based representation of a structured and interpretable latent space built from multimodal neuroimaging data, aimed at capturing disease heterogeneity and facilitating subtype discovery across both AD and FTD.

We analyzed 1,522 18F-FDG PET and MRI scans, split into training (60%), validation (20%) and test (20%) sets, stratified by diagnosis and age to avoid subject overlap. The model was trained using data from 336 cognitively normal (CN), 504 AD and 106 FTD subjects. We extended the DIVA framework [6] to a multimodal setting, integrating auxiliary tasks for diagnosis and age prediction, along with latent space consistency constraints. A graph was constructed from the latent space using a dissimilarity measure derived from latent consistency. PatientSpace was evaluated through (1) agglomerative clustering over the graph and (2) local neighborhood analysis on an unseen test set. Additionally, longitudinal scans from 673 mild cognitive impairment (MCI) subjects were projected into the latent space to study dementia conversion.

PatientSpace effectively distinguished CN, AD and FTD groups, achieving classification performance comparable to state-of-the-art approaches : CN (Sensitivity: 0.99 - Recall: 0.87); AD (Sensitivity : 0.91; Recall : 0.92) and FTD (Sensitivity 0.70; Recall : 0.91). Ten distinct clusters were identified, differing in terms of diagnosis, age, MMSE, atrophy and metabolism patterns. Five AD clusters have been identified. All exhibited typical patterns of brain atrophy and metabolic changes, except for two clusters – one showing only atrophy without distinct metabolic pattern, and the other lacking any clear pattern when compared to CN groups. Additionally, two FTD clusters emerged. One aligned with the classical FTD profile, displaying frontal atrophy and hypometabolism, while the other presents a less typical pattern, characterized by temporal atrophy and hypometabolism without involvement of the frontal regions. Neighborhood analysis demonstrated coherent spatial organization with similar subjects sharing neuroimaging features. MCI projections revealed distinct progression trajectories, with most subjects initially positioned closer to CN clusters, and early separation between converters and non-converters observable at baseline.

PatientSpace provided a unified representation of neurodegenerative diseases, capturing both global and local heterogeneity. While the framework successfully differentiates AD, FTD and CN populations, limitations include class imbalance, potentially affecting FTD classification performance, and modality-related sensitivity differences, as PET may detect earlier changes than MRI [7]. Discrepancies with previously reported AD subtypes may stem from differences in input modalities, model strategies and cohort composition [8], [9]. MCI projections supported the model’s potential to detect early conversion risk through multimodal embedding.

We present PatientSpace, a graph-based latent space framework that enables interpretable, multimodal representation learning for the study of AD and FTD. By capturing subject-level heterogeneity and disease trajectories, PatientSpace offers new opportunities for data-driven subtyping and personalized analysis of neurodegeneration.

Progressive Supranuclear Palsy (PSP) is diagnosed in the presence of either supranuclear gaze palsy or a combination of slow vertical saccades and postural instability with falls within the first year1. On the same way, postural instability is one of the cardinal signs in Parkinson’s Disease, it is associated with increased falls and loss of independence2. For this reason, PSP and PD, especially in their early stages, show overlapping clinical manifestations3. The aim of the study was to identify the mainly brain regions involved in both diseases to show the differences between PD and PSP diseases using machine learning approach.

In this study, we enrolled 115 PSP from the Neuroscience research center at Magna Graecia University of Catanzaro and 100 PD from the Parkinson’s Progression Markers Initiative (PPMI). All patients underwent a 3T brain MRI scan. We utilized FreeSurfer version 7.1.1 software for automated neuroanatomical segmentation on T1-weighted MRI scans. The "recon-all" function was used to smooth and normalize the images to the "fsaverage" template, reshaping them into an inflated surface. Then, the Desikan-Killiany atlas was applied to divide the cortex into 34 regions of interest (ROIs) per hemisphere. Several morphometric metrics were extracted for each ROI, including cortical thickness (CT), cortical and subcorticall volumes (SV), mean curvature (MC), and estimated total intracranial volume (eTIV). We split the dataset into training and test sets with an 80%–20% ratio. A machine learning (ML) algorithm based on decision trees, specifically XGBoost4, was applied to the entire dataset. Our aim was to differentiate individuals with Parkinson’s disease (PD) from Progressive Supranuclear Palsy (PSP) using morphometric data. To evaluate the model’s performance, we employed nested k-fold cross-validation (n-CV), which consisted of two loops: an outer loop with 4 folds and an inner loop with 4 folds. In the inner loop, we fine-tuned the algorithm and selected the best parameters. The outer loop then estimated the overall performance of the selected model. The final performance metrics were averaged across the different algorithms. The goal of nested CV is to include an algorithm selection step within the inner loop, so any final model to be trained or deployed must go through the same process used in the inner loop (see Figure 1). Our feature selection approach involved two main steps. First, we evaluated the collinearity of features using correlation analysis, removing any features with a Pearson correlation coefficient greater than 0.6. Next, we trained XGBoost models on the remaining features in each inner fold after hyperparameter tuning. To maximize accuracy, we performed a randomized search (with ten iterations) to optimize parameters like maximum depth, minimum child weight, gamma, number of estimators, subsample, and learning rate. Continuing with feature selection, we assessed feature importance using SHapley Additive exPlanations5 (SHAP). SHAP is a method that explains each prediction by quantifying the contribution of each feature to the model’s output, approximating Shapley values from game theory. Specifically, the SHAP value represents a feature’s contribution to the difference between the actual prediction and the average prediction. In our pipeline, features were ranked based on SHAP importance, and we iteratively removed them to improve accuracy metrics. The performance of the tuned XGBoost model was evaluated based on the mean area under the curve (AUC), accuracy, sensitivity, and specificity in the validation folds of the outer loop. Finally, the best model was tested on the independent test set created initially and on the entire training dataset.

Before applying feature selection, the XGBoost classifier using the nested cross-validation method achieved an accuracy of 0.73 and an AUC of 0.80 between PD and PSP patients. After implementing feature selection using SHAP, resulted in an accuracy of 0.82 (AUC 0.87) with 11 selected features (Figure 2). Figure 3 provides useful insights into how specific brain structural measurements impact model predictions. Features such as the left accumbens, left cerebellum white matter, some frontal and temporal thickness, 2 temporal mean curvature and 2 frontal areas had significant effect.

This pilot study showed the important difference between PSP and PD, the importance of the results is attributed to the robustness of the nested cross validation method. Another significant aspect of the study is the crucial role of correlation, considering the interrelationship noted among morphometric characteristics. The capability to integrate parameter tuning into the optimization of machine learning models enhances model accuracy while minimizing the risk of overfitting6.

Our study aids in the progression of neurodegenerative disease research and emphasizes the significance of using innovative computational approaches to enhance patient care outcomes.

Accurate segmentation of brain tumors from Magnetic Resonance Imaging (MRI) scans is crucial for diagnosis, monitoring disease progression, and formulating treatment strategies. While traditional methods (RANO & iRANO) are labour-intensive, advanced deep-learning techniques are being explored for more efficient and precise segmentation. In this study, we compare the performance of six state-of-the-art segmentation models: U-Net, nnU-Net, Attention-UNet, ELU-Net, U-Net++ and UNETR [1-7], using the BraTS [8-9] and a private experimental dataset consisting of T2-FLAIR (Fluid-Attenuated Inversion Recovery) and Post-Contrast T1-weighted MRI scans. We evaluate the model performances based on Dice and Jaccard scores [10-11], with nnU-Net consistently achieving the highest accuracy, particularly after transfer-learning on the experimental dataset. To address performance degradation in cross-domain image scenarios, we introduce MAMU-Net by enhancing the base nnU-Net architecture with a novel Multi-Activation Module that fine-tunes the activation layers.

The BraTS datasets contain MRI scans (from the year 2019-2021) from 19 institutions stored in NIfTI format. We've used post-contrast T1-weighted (T1PC) and T2-FLAIR modalities for training different UNet models [Fig. 1] using the hyper parameters as listed in Fig. 1. Expert neuro-radiologists validated manual annotations, marking three regions: Necrotic and non-enhancing tumor core, Enhancing tumor region, and Peritumoral edema. A private MRI dataset of 64 glioblastoma (GBM) patients with 2,413 slices was acquired at the University of Pennsylvania under a collaboration with TCG CREST (data sharing ID: RIS76150) using a 3T Tim Trio MR scanner. High-resolution post-contrast T1-MPRAGE and T2-FLAIR images were coregistered to lower-resolution Diffusion Tensor Imaging (DTI) - B0 images, with a resolution of 128x128x35, as part of a continuation of a project focused on quantitative MR imaging, resulting in lower-resolution images being used for segmentation. Using a semi-automated algorithm, lesions were segmented into Contrast-enhancing (ET), Non-enhancing (NCR), and Edema [12]. This is used as a ground truth. Manual skull stripping was necessary as automated tools struggled with the lower-resolution images. For model optimization, we use Binary Cross-Entropy as the loss function. We evaluate the model's performance using the Dice score and Jaccard Index [13-16]. We aimed to study the impact of different variants of linear unit-based activation functions, i.e. “CELU”, “ELU”, “GELU”, “LeakyReLU”, “PReLU”, “ReLU”, “ReLU6”, “RReLU”, “SELU”, “SiLU”, [17-27] on the best model among “UNet”, “nnUNet”, “Attention UNet”, “ELUNet”, and “UNet++”. The purpose is to enhance performance further. Fig. 3 provides brief details about the activation functions used in this study. We trained a single model in which the activation function is represented as a weighted sum of the ten activation functions described above. We extracted the weights assigned to each activation function and computed their normalized values for each layer. We ensured that each normalized value falls within the range of 0 to 100, and that the sum of all normalized values in each layer equals 100. We derived our proposed model “Multi-Activation Module Unet (MAMU-Net)”, by modifying the activation function layer of “nnU-Net” and introducing a new Multi-Activation Module in its architecture. The “Multi-Activation Module” is linear combination of “PreLU” and “ReLU6”. Here the coefficients of the linear combination are also trainable parameters.

On both datasets, i.e. BraTS validation dataset and the experimental datasets, nnUNet performed best with a Dice score of 88.95% and 64.07% respectively [Fig. 2]. However, the performance of our proposed MAMU-Net was closely aligned with that of nnU-Net. But after transfer learning of last stage, MAMU-Net achieved 74.64% Dice score almost 5% higher than nnUNet (69.86%) [Fig. 2]. After retraining all the decoder stages, MAMU-Net remained best with 85.65% Dice score.

Most available open-source datasets consist of pre-operative MRI scans taken before surgery. In contrast, our study evaluates treatment response using post-operative MRI scans, which often include surgical cavities. This poses a challenge in segmentation due to the midline shifting of brain. However, in all types of data nnUNet remained the top performer achieving a Dice score of 88.95% and a Jaccard score of 80.43%. However, it lacked the flexibility to be applied across multiple data sources. Our newly introduced MAMU-NET, featuring the novel Multi-Activation Module, has demonstrated faster feature capture on a new dataset compared to nnU-Net. This makes MAMU-Net more adaptable for domain-specific feature learning. Retraining experiments showed that MAMU-Net achieved the most significant performance improvement, reaching a Dice score of 74.64%, which is 5% higher than that of nnU-Net.

Conventional parametric methods for MRS metabolite quantification suffer from major drawbacks such as modeling biases due to the variability in how prior knowledge, constraints, and parameters are accounted in the fit. Deep learning methods [1,2,3,4] have also been explored. As most of them rely on a training on simulated data, they might also suffer from modeling biases but this also raise the question of a domain shift when used on real data. A question that remains unaddressed in the context of MRS. Domain adaptation refers to the challenge of using a model on a test distribution different from the training distribution. Domain adaptation (DA) methods can be either feature based [5,6], data based [7,8] or model based [9,10]. In this work, we propose an experimental protocol to highlight the domain adaptation problem in the context of deep learning for MRS quantification, show which MRS parameters can be a source of domain shift, and then test the efficiency of different DA methods in improving transfer across MRS domains. In this work, we propose an experimental protocol to highlight the domain adaptation problem in the context of deep learning for MRS quantification, show which MRS parameters can be a source of domain shift, and then test the efficiency of different domain adaptation methods in improving transfer across MRS domains.

We use a simulator based on the Eq. 1, to generate datasets of short echo time (TE=30ms PRESS) MR spectra by randomly sampling values for the amplitude, damping, frequency for each metabolite (simulated with the GAMMA library) and a phase. Each dataset corresponds to a different domain. We generated 8 different databases (A, B, C… H) of 200000 samples each by combining two different spectral line shapes (lorentzian or gaussian) and four different configurations for macro-molecule profiles. We model the macro-molecule signals as a sum of eleven different macromolecular components, with a base amplitude each, and a variation percentage. The characteristics of each dataset are detailed in Table 1. For each DA method, we trained one model per domain and tested it on every other domain. A domain shift matrix was then computed to measure the quantification error of each model when it is used each domain. Our goal is to find DA methods to make models perform as well on others domains as on the training domain: we want the domain shift matrix to be as close to zero as possible. Our quantification model is a convolutional neural network, trained in a supervised fashion. It is composed of two parts: an encoder composed of four convolutional layers, each followed by a batch normalization layer, an activation function and a sub-sampling, and a regression head composed of two fully connected layers. The output of the model for a spectra is a vector composed of the amplitude, damping factor and frequency shift of each metabolite and a global phase. Our baseline is this model without any DA method. We then combine this model with the test-time batch normalization (TTBN) [11]. This method replaces the mean and standard deviation computed on the training dataset of each BatchNorm layer by those recomputed on the test dataset.

Values in the domain shift matrix are the logarithm of the MSE of the amplitude between the predicted values and the ground truth. Each line correspond to a training domain and each column to a test domain. The matrix was normalized by subtracting the diagonal (same train and test domain) to each line. In fig. 2, the matrices "Base model" and "TTBN" respectively show the quantification errors for the base model and the TTBN.

For the base models, best results are mainly obtained on the diagonals. This is the expected result, as a model should indeed perform better when applied on data similar to those that it was trained on. Performance is degraded in other cases. Models A, B, E and F, that were trained with lower macro-molecule amplitude variations, perform worse when tested on domains with different macro-molecule base amplitudes, while models C, D, G and H appear to suffer less from the domain shifts. Regarding the TTBN method, even though the domain shift metric is still mostly the lowest on the diagonals, the matrix is more homogeneous, and the performance on other domains are less degraded. This shows that the TTBN method improves indeed transfer across domains.

First, one can see that the performances of deep neural networks for MRS quantification can be altered by a domain shift: the macromolecular profiles and the spectral line shapes can be a source of domain shift between databases, but other parameters have to be considered, and might also be a source of data distribution shift. Second, we show the importance of considering DA methods. TTBN appeared to be efficient in improving transfer across domains. Other DA methods from the literature should also be tested, and the design of DA methods that would suit the specificities of MRS signals should also be considered.

The widespread use of neural networks for MRI data analysis and the limited availability of real data for training underscore the need for synthetic MRI datasets [1, 2]. To generate these synthetic datasets, digital phantoms are essential. A digital MRI phantom (in the basic version) comprises tissue characteristics such as T1 and T2 relaxation times and proton density (PD) [3]. These phantoms simulate the properties of various tissues during MRI acquisition [4]. We aim to investigate the capabilities of neural networks, particularly diffusion models, for generating digital MRI phantoms from weighted MR images. The novelty of our study lies in the simultaneous generation of three quantitative maps: T1, T2, and PD.

A) Dataset. The training dataset consists of quantitative maps of T1, T2 relaxation times, and proton density (PD), along with corresponding synthetic MR images weighted for T1, T2, and PD. Thus, the dataset contains 3482 pairs of quantitative maps and weighted images. These quantitative maps were generated using the methodology outlined in [5], employing anatomical brain models from the BrainWeb database [3]. These models include 11 distinct tissue types, each assigned specific T1, T2, and PD values sourced from the literature [4, 6, 7]. To introduce variability and prevent uniformity, tissue parameter values within each anatomical region were sampled from a normal distribution centered around the mean values, with a standard deviation of 10%. To obtain synthetic weighted images, we developed a pipeline, illustrated in Figure 1. The process involves three key steps: generating digital phantoms from quantitative maps (T1, T2, PD), simulating MR signals using Bloch’s equations (via the KomaMRI software package [8]), and reconstructing the final MR images. In this study, employed pulse sequences were T1-, T2- and PD-weighted 2D turbo spin echo (TSE) (TR = 700ms/3.0s/6.0s, effective TE = 7.65ms/76.5ms/6.69s, ETL =10, resolution – 1.95*1.95*5mm^3, field-of-view – 0.25*0.25m), constructed using the open-source Pulseq framework [9]. KomaMRI also requires T2* maps for input, in our case they were replaced with T2 maps, because sequences based on TSE are generally not sensitive to T2* effects. B) Model. Denoising Diffusion Probabilistic Models (DDPMs) excel in generating high-fidelity images and are widely used in computer vision [10]. However, their reliance on extensive diffusion steps results in prolonged training and sampling times [11]. To address this, we adopted the Fast Denoising Diffusion Probabilistic Model (Fast-DDPM), which achieves comparable image quality with fewer than 1000 time steps [11]. During training, the model uses quantitative T1, T2, and PD maps as inputs, with T1-, T2-, and PD-weighted MR images serving as conditioning data. At inference the model takes weighted MR images (as the condition) and Gaussian noise (as input) to predict the underlying quantitative maps. Our implementation used 800 diffusion steps with a uniform noise scheduler. Training was conducted over 763 epochs with a batch size of 8, and the image size was 128 × 128 pixels. Figure 2 illustrates the training and inference process of the model. C) Evaluation. The test set consisted of 100 brain slices from a separate anatomical model (T1, T2, PD-weighted images). To evaluate the model's quality, metrics such as the Structure Similarity Index (SSIM) [12] and Peak Signal-to-Noise Ratio (PSNR) [13] were used.

The trained Fast-DDPM successfully generated quantitative T1, T2, and PD maps. To assess their quality, Figure 3 provides a direct comparison between the ground truth and synthetic maps. Figure 4 shows the probability density curves for the original and synthesized maps. For the T1 quantitative maps SSIM achieved 0.78±0.05, and PSNR - 22.12±0.55 (average of 100 slices). For T2 quantitative maps SSIM – 0.79±0.02, PSNR – 18.67±0.45, for PD quantitative maps SSIM – 0.55±0.04, PSNR – 19.32±0.35.

The synthetic phantoms produced by our method exhibit high structural similarity to the original ground truth for T1 and T2 maps, and moderate for PD maps. Notably, our framework simultaneously generates all three quantitative maps (T1, T2, and PD) without relying on prior segmentation. To date, we have been testing the model on synthetic data from a separate anatomical model. In the future, we plan to test the model on real weighted MR images and real ground truth quantitative maps.

The DDPM-based approach shows potential for creating realistic digital phantoms. The creation of realistic brain phantoms holds significant promise for advancing MRI research - for pulse sequence and reconstruction algorithms design, and for generation of synthetic MRI datasets. Acknowledgements This work was supported by the Ministry of Science and Higher Education of the Russian Federation (Project FSER- 2025- 0009).

The Choroid Plexus (ChP) is a vascular brain structure involved in vital brain functions [1,2]. Most clinical studies focused on analyzing ChP Volume (ChPV) in diseased populations, hypothesizing a link between increased ChPV and neuroinflammatory processes, although this association is influenced by multiple factors [3,4]. ASCHOPLEX [5], a deep learning-based toolbox with an integrated fine-tuning step, returns accurate ChP segmentations on non-contrast-enhanced T1-weighted MRI images [6], but struggles with data variability [7]. To enhance its robustness, this work integrates ASCHOPLEX into Dafne [8], a federated incremental learning framework that supports privacy-preserving model updates [9,10]. In contrast to classical fine tuning, Dafne enables adaptation to new data while retaining previously acquired knowledge by performing a weighted averaging of the pre-existing and the fine-tuned model at every aggregation. This study aims to evaluate the generalizability of ASCHOPLEX by comparing the performance of fine-tuning and federated learning strategies across two Multiple Sclerosis (MS) cohorts.

Details of the data are provided in Tab.1. Dataset 1 (D1) consists of 128 subjects acquired at the Multiple Sclerosis Centre of the University Hospital of Verona, Italy [5]. Dataset 2 (D2) includes 19 MS subjects from the publicly available ISBI 2015 Longitudinal MS Lesion Segmentation Challenge, including the first time point for each subject [11,12]. ASCHOPLEX architectures were modified by incorporating the following models implemented in MONAI [13]: DynUnet [14], UNETR [15], and SwinUNETR [16]. A five-fold cross-validation training was performed on D1. Training parameters and data augmentation techniques were consistent with those described in [5], with the addition of data augmentation transforms from TorchIO [17]. The five best-performing fold configurations were selected based on the Dice Coefficient and combined using an ensemble majority voting strategy. These models were then separately fine-tuned following the same procedure as in [5], and integrated into Dafne [8] to enable incremental learning on D2 using 400 epochs. Dafne was adapted to support 3D models in MONAI, including ensemble processing and server-side model merging. Three prediction strategies were evaluated on the test sets of both datasets: direct prediction following initial training on D1 (DP), prediction after fine-tuning on D2 (FT), and prediction following incremental learning on D2 (IL). The predicted segmentations were compared to expert-generated ground truth (GT) manual segmentations. Segmentation performance was assessed using absolute ChPV, Dice Coefficient, 95th percentile Hausdorff Distance (95% HD), and Percentage Volume Difference (ΔVol%), using Python (v 3.12).

Selected configurations include two UNETR models, a DynUNet model, and two SwinUNETR models with a patch size of 128 and loss function combining Generalized Dice and Cross-Entropy. The segmentation metrics (Tab.2, Fig.1) on DP report a drop in performance on D2 compared to D1 (mean Dice/ΔVol%: D1=0.80/8.96%, D2=0.35/-71.53%). In contrast, FT and IL approaches on D2 show improvements compared to DP (mean Dice: FT=0.56, IL=0.43), despite ΔVol% consistently exceeding 55%. Concerning the retesting on D1 after ASCHOPLEX’s exposure to D2, FT performance drops significantly compared to DP (mean Dice/ΔVol%: 0.12/-89.02%), with 95%HD considered unreliable. Meanwhile, the IL approach yields ChPV estimates that are closer to DP (Δ(ΔVol%): IL-DP= -1.8%, FT-DP=-98%). Fig.2 displays the compared segmentations overlaid on the T1-weighted MRI for a representative subject from both D1 and D2.

This study investigates the generalizability of fine-tuning and federated incremental learning approaches for the ChP segmentation task. The findings indicate that fine-tuning deep learning models on a new dataset enhances performance compared to direct prediction using a model trained on a different dataset. However, this improvement comes at the cost of an irreversible loss of information from the dataset used for initial training, mostly when datasets have different image intensity range (Fig.2), as demonstrated by the performance degradation observed when re-evaluating the fine-tuned model on the D1 test set. In contrast, the federated incremental learning strategy implemented within the Dafne framework yields comparatively lower performance but offers greater generalizability. Specifically, merging the model trained on D1 with a model briefly trained on D2 improves prediction accuracy on D2 relative to direct prediction, while preserving the knowledge previously acquired from D1.

To conclude, integrating ASCHOPLEX into the federated incremental learning Dafne framework enables better generalization across datasets by preserving prior knowledge, offering a robust alternative to fine-tuning for ChP segmentation in MS studies.

Multi-echo turbo-spin echo (ME-TSE) is a fast MRI pulse sequence used for T2 mapping, offering significant speed advantages over alternatives like multi-echo spin echo (MESE). This efficiency is critical for breath-hold scans or dynamic studies. Dictionary-based T2 mapping (Echo Modulation Curve - EMC method) has proven effective for ME-TSE and MESE data, but is computationally intensive [1, 2, 3], as it requires direct matching of pixel curves and the entire dictionary. The original EMC method authors reported a reduction in fitting time by implementing a gradient descent (GD) optimization scheme [4]. An alternative acceleration strategy combines neural networks (NN) with the EMC method. For instance, a U-Net trained on brain MESE data successfully replicated EMC-derived T2 maps [5]. Shifting from image-based to dictionary-based training could improve robustness across different scan areas, as shown for upper leg T2 mapping [6]. Yet, that approach simulated echo modulation curves using extended phase graph (EPG) formalism, which does not account for slice-profile effects. Moreover, EPG’s computational complexity escalates exponentially with longer pulse sequences. While neural networks have been applied to MESE, their adaptation to ME-TSE, where echo modulation curves typically comprise only ~3 values (vs. many in MESE), remains unexplored. Here, we investigate a neural network-based T2 mapping method for ME-TSE, trained on a dictionary of Bloch-simulated echo modulation curves.

Agarose-based phantoms (Eurospin) with calibrated reference values (T1ref= 336.1/672.5/1235.2/1288.9/1407.7 ms and T2ref= 97.2/118.5/170.5/158.7/131.6 ms) were scanned at 3T (Magnetom Vida, Siemens). The ME-TSE pulse sequence parameters were: TR = 1.5 s, first TE = 7.4 ms, inter-echo time = 7.4 ms, turbo factor = 3, number of contrasts = 3. Additionally, an in vivo scan of the lower leg muscle was acquired using identical parameters, except for TR = 2 s. In ME-TSE, k-space for each image is filled with data acquired at different echo times, where the number of such echoes is defined as the turbo factor. We employed an effective TE combination identified in our previous work [7]. Using a Matlab script [3], five EMC dictionaries were generated for each phantom with parameters: B1+ = [80:1:120] %; T2 = [50:1:240] ms; T1 = T1ref of phantom. For muscle one EMC dictionary were generated with next parameters: T2 = [5:0.25:100] ms; T1 = 1150 ms. The dictionaries were augmented with Rician noise based on SNR characteristics estimated from MR images. Training data size after augmentation for each phantom: 15662 curves, for muscle: 31242 curves. Similar to [6], a fully connected neural network (multilayer perceptron) was selected using the Keras library. The training and inference pipeline for the neural network are presented in Figure 1. Initially, T2 maps were reconstructed using the EMC dictionary method. After that, T2 maps were generated based on the neural network illustrated in Figure 1.

In Figure 2A, T2 maps obtained using various methods are presented. Visually, the maps produced by the EMC method and the NN method appear similar; however, the T2 maps for the NN method exhibit less noise. Figure 2B shows a correlation plot between the T2 values of each pixel inside region of interest for both methods, with calculated Pearson coefficient. Figure 2C displays the mean relative errors compared to the reference T2 values for both methods. It can be seen that both methods have approximately the same error, however, NN-method demonstrates lower standard deviation of the error. The T2 maps for muscle tissue obtained using the two different methods also show no significant difference (Figure 3A), as can be seen in the difference map of the two (Figure 3B). Figure 3C presents a correlation plot for T2 values of each pixel in five regions of interest for the muscle (Figure 3D). A high correlation with a strong Pearson coefficient is observed.

In this study, we demonstrate the feasibility of a T2 mapping method based on neural network trained on simulated echo modulation curves obtained using Bloch simulations for a multi-spin turbo spin-echo pulse sequence. A high correlation between the results of using NN and EMC methods were observed. Additionally, the NN fitting took nearly 13 times less time than the classic EMC method, requiring only 1s compared to 12.87s for a 66x197 matrix. In the future, the NN method could be used for multi-slice T2 mapping and could also be adapted to estimate another parameters, such as T1.

In conclusion, it has been demonstrated that the use of neural networks can effectively replace the EMC dictionary-matching method, resulting in lower T2 estimation errors and higher processing speeds. By employing Rician noise augmentation and training the data with low T1-weighting, we achieved a very low median relative error in T2 estimation. This study was supported by the Russian Science Foundation (RSF) grant No. 23-75-10045.

Personalized spinal cord models are essential for enhancing the effectiveness of epidural electrical stimulation (EES) therapies aimed at restoring neurological functions in individuals with spinal cord injury (SCI)[1–6], and other diseases [7, 8]. Built on detailed, patient-specific medical imaging data, these models play a critical role in both the pre-operative planning and post-operative fine-tuning of EES implant placements. The processing and annotation of medical imaging required to create the spinal cord personalized model is labor-intensive and time-consuming[3]. Anatomical tissue segmentation is a bottleneck for the scalability and widespread clinical adoption of personalized spinal cord models. Artificial neural networks (ANNs) hold promise for automating tissue segmentation from medical imaging, thereby reducing reliance on manual intervention. ANNs have proved successful in many segmentation tasks[9]. However, they face challenges due to the scarcity of annotated medical data and the variability between patients, imaging protocols, anatomical regions, and scanner types[10, 11]. Self-supervised learning (SSL) has emerged as a powerful method for deriving meaningful representations without human supervision, leveraging available unlabeled data. SSL has fueled the success of natural language processing (NLP) models like ChatGPT[12], which learn the structure and semantics underlying language by predicting masked (hidden) elements in sequences of words based on the surrounding context. This self-supervised objective enables NLP models to perform downstream tasks such as text summarization with remarkable efficiency. Taking inspiration from NLP, we applied SSL methodologies to unlabeled spinal MRI volumes to develop generalistic models that can be adapted to downstream tasks with minimal expert intervention. Our framework, SpineGPT, learns by reconstructing masked parts of the MRI volumes from their visible counterparts, extracting representations that capture the spinal cord’s underlying structure. We use these representations to fine-tune SpineGPT to solve clinically-relevant problems such as tissue segmentation.

Our current implementation of SpineGPT builds on a masked auto-encoder (MAE) framework[13], learning self-supervised representations of MRI scans through a self-reconstruction objective. In this process, a 3D chunk of the scan is largely masked (80%) at random and used as input to an autoencoder neural network, which is trained to reconstruct the missing portions from the visible parts (Figure 1). Through this pretraining phase, the model’s encoder learns to map the original volume to a latent representation that captures informative structural details. SpineGPT’s current encoder and decoder consist of standard Vision Transformer (ViT) layers adapted for 3D volumes. We leverage all available unlabeled data, totaling over 1,300 diverse T2 weighted MRI volumes of the spine from both publicly available[10, 14–17] and private datasets[3, 4, 6–8, 18]. Segmenting spinal tissues such as spinal roots, white matter (WM), and cerebrospinal fluid (CSF), is vital for developing personalized spinal cord models. Current methods, based on supervised learning, allow for segmentations for tissues except spinal roots in the lumbosacral region[19–23]. We propose a strategy of freezing the encoder of SpineGPT and finetune the decoder to segment previously described spinal tissues.

The input volumes are cropped around the spinal cord[21, 24] into chunks of 32x32x64 with patch sizes of 4x4x8. The current architecture of the encoder of SpineGPT consists of 12 ViT layers with 16 self-attention heads[25, 26]. The decoder consists of 8 ViT layers. The pre-trained phase minimizes the reconstruction error between the input patches and the original patches using Mean Square Error (MSE). We finetuned the decoder on the available labeled datasets containing CSF, WM and spinal roots [3, 6–8] consisting of 31 individual T2 weighted high resolution MRI (Resolution 0.3x0.3x0.6 mm3). We minimized the Dice loss[27]. We tested our segmentation framework on two test sets. First, we tested on a subset (2 individuals) containing both dorsal and ventral spinal roots. Secondly, we tested on a public labeled data set of 14 participants on the lumbosacral region[28] only containing dorsal roots. We achieved promising preliminary results, with Dice scores of 0.95±0.01, 0.94±0.01 and 0.73±0.00 on the first test set (Figure 2) and 0.91±0.03, 0.89±0.23 and 0.41±0.17 on the second test set for segmentations of CSF, WM and spinal roots respectively.

We introduce SpineGPT, a foundation framework for spinal cord T2 weighted MRI. We finetuned SpineGPT to address, for the moment, one downstream task, lumbosacral spinal tissues segmentation, in particular spinal roots. Preliminary results are promising, achieving state-of-the-art results in the segmentation task. However, extensive testing is still necessary to ensure a robust framework.

Cardiac MRI T1 mapping quantifies the T1 relaxation time of cardiac tissues, providing critical insights into tissue composition and aiding in diagnosing pathologies such as fibrosis, and diffuse myocardial inflammation. Data acquisition requires multiple breath-holds and is sensitive to motion artifacts, limiting feasibility for high-resolution imaging. Compressed Sensing has shown success in accelerating data acquisition, however most existing methods do not consider the physical exponential decay model in optimization, and rely on fixed undersampling patterns that do not fully exploit acceleration potential.

Our pipeline is composed of 3 conceptual layers (see attached figure) - first, a subsamlping layer, that models a learned, per-frame parametric non-Cartesian acquisition set,and applies it over the input data. Second, a neural reconstruction block, performing denoising before decay estimation, to better-inform both the decay estimation model in the forward pass and the samplng scheme in the backward pass. Lastly, we use a decay estimation model, which a neural network that outputs exponential decay parameters. These parameters are used to perform exponential regression, and the residuals of this regression are used as a loss objective that is backpropagated through all 3 layers. For acquisition parameterization we follow [1], for the reconstruction and decay models we use the reconstruction model from [2]. To efficiently leverage data-priors for learning an acquisition set that would generalize well across multiple sequences, we first optimize a non-Cartesian acquisition set solely for reconstruction (train only the sampling and reconstruction layers). At a second stage, we finetune this set for decay estimation, guided by the physical decay model. Finally, we propose finetuning the reconstruction and decay models per-sample for optimal estimation results.

The core of our work is showing the learned, per-frame learned acquisition schemes, alongside the inclusion of the physical decay model in optimization, greatly benefit T1 Mapping acceleration. To show this, we compare T1-PILOT to four baselines that are representative of current approaches for acquisition and learning objectives used in training – radial acquisition (constant sampling scheme shared across frames) , GAR acquisition (constant per-frame sampling), a single, learned acquisition pattern shared across frames, and per-frame learned acquisition not trained under the physical decay model. T1-PILOT significantly outperformed baseline strategies, achieving higher T1 map fidelity at greater acceleration factors. Specifically, it demonstrated consistent gains in Peak Signal-to-Noise Ratio (PSNR) and Visual Information Fidelity (VIF) relative to existing methods, along with marked improvements in delineating finer myocardial structures.

We demonstrate the advantage of incorporating physically feasible, per-frame non-Cartesian learned acquisition schemes in T1-Mapping acceleration. We also show that intergrating the physical decay model directly within the learning objective dramatically improves acceleration potential. This approach addresses the limitations of traditional methods that rely on fixed undersampling patterns and demonstrates the potential of physics-informed deep learning in cardiac MRI. We hope these findings would encourage and inform future endeavors in T1-Mapping and general MRI acceleration. One limitation of our work is that the multiple optimization steps require higher computational resources, in comparison with current methods. We aim to explore alternative optimization approaches that would alleviate this difficulty. Another limitation is that our finidngs mostly rely on simulated data created from fully-sampled k-space. We aim to test our method on actual MR machines in the future.

The integration of the T1 signal relaxation model into the sampling-reconstruction framework allows T1-PILOT to optimize k-space trajectories effectively, leading to enhanced quantitative accuracy and reduced acquisition times. This approach addresses the limitations of traditional methods that rely on fixed undersampling patterns and demonstrates the potential of physics-informed deep learning in cardiac MRI.

Cancer remains the primary cause of mortality globally [1], underscoring the need for leading-edge diagnostic tools. Magnetic Resonance Imaging (MRI) is a non-invasive, frequently used modality in brain tumor detection as it provides insights of clinical importance, ensuring early brain tumor detection [2]. MRI offers precise identification of tumors by contrasting soft tissue without ionizing radiation [3]. Machine learning based diagnosis systems have demonstrated enhanced diagnosis and classification results by utilizing advanced CNNs (Convolutional Neural Networks) for MR images. For instance, Residual fused shepherd CNN [4] is a hybrid model that provided accuracy of 94% in brain tumor detection. In [5], Binary classification of brain tumors has been explored using several CNN variants, presenting the performance analysis of Alexnet, VGG-16, GoogleNet and RNN. The evaluation on BRATS 2013, BRATS 2015, and OPEN I dataset revealed that AlexNet achieved the highest accuracy of 98.7%. Based upon the insights from literature, it is evident that CNNs has a very crucial rule for brain tumor detection in MR images. A most important advancement in CNN optimization is EfficientNet [6], which employs compound scaling to balance network depth, width, and resolution, to improve model efficiency and accuracy. Therefore, it is worthwhile to analyze its impact in comparison with conventional architectures. The main objective of this study is to analyze the impact of EfficientNet’s compound scaling method on conventional VGG-16, VGG-19 [7], ResNet-50, ResNet-101 [8], and DenseNet-121, DenseNet-169 architectures [9].

In this study, the compound scaling strategy of EfficientNet is applied to each of the conventional architecture as explained in the introduction section and the performance of unscaled conventional models is compared with their respective scaled versions. Firstly, a customized dataset is made by combining the datasets created from different sources comprising MR scans, that is Figshare [10] and Kaggle sources [11],[12]. The dataset is categorized into four tumor classes (glioma, meningioma, pituitary tumor and no tumor) and preprocessed through resizing (130×130 for conventional and 156×156 for enhanced compound scaled variants) followed by normalization. The EfficientNet model width scaling in conventional CNN is represented by Eq.1: w′ = w * β^∅  (Eq. 1) The resolution scaling is represented by Eq.2: r′= r * γ^∅   (Eq. 2) Where parameters in above equations are; w: original width (number of filters), r: original input resolution, β: width scaling factor, γ: resolution scaling factor, ∅: compound coefficient, w′, r′: scaled width and resolution, respectively The value of compound coefficient ∅ is set to 1.2 as explained in EfficientNetB2 [6]. The compound scaling is applied in width and resolution only as shown in Figure 1. The depth of conventional CNNs remains constant as each architecture is designed with a predefined depth factor. All models are trained with 80-20 training-testing split, using categorical cross-entropy loss function and Adam optimizer. Performance metrics include accuracy, precision, recall, and F1-score.

The impact of EfficientNet compound scaling on CNN architectures is evaluated using key performance metrics including accuracy, precision, recall and F1-score. Table 1 shows the model comparisons and highlights the impact of compound scaling on different CNNs architecture performance. The best accuracy results were achieved by DenseNet-121(90.21% unscaled, 95.81% scaled), followed by scaled VGG-16 and VGG-19 at 94.68%. ResNet-101 exhibited a substantial accuracy gain, rising from 73.24% to 90.84% after scaling. ResNet-50 demonstrated relatively low accuracy (77.13% unscaled, 81.83% scaled).

The analysis of proposed scheme across all models shows that the most prominent improvement from compound scaling is observed in ResNet-101, with accuracy increasing by over 17%. Overall, the proposed strategy for scaling the conventional models demonstrates better generalization and fast convergence. Figure 2 demonstrates the performance of each model in classifying brain MR images for tumor predictions, under unscaled and scaled conditions. All the models have demonstrated strong performance in tumor classification after applying compound scaling. The future research could focus on refining scaling parameters and exploring more effective data augmentation techniques to further enhance model performance.

The results clearly indicate that the EfficientNet compound scaling strategy significantly improved accuracy of conventional models. This highlights the substantial impact of architecture choice and scaling on the performance of conventional CNNs in MRI based brain tumor detection. Although the effectiveness of this method varied across models, the EfficientNet scaling method enhances the feature learning and overall classification performance.

Treatment planning and surveillance of patients with glioma can benefit greatly from accurate tumor segmentation [1,2]. However, manual segmentation is time-consuming and subject to intra- and inter-rater variability [3]. While recent advances in artificial intelligence (AI) have yielded automated segmentation tools with expert-level performance [4], many existing tools rely on specific preprocessing steps or consist of separate installable components, complicating their use in clinical research settings [5]. We present Glioseg, a fully automated, end-to-end framework for multi-compartment glioma segmentation that integrates robust preprocessing, multiple AI models, and ensemble-based label fusion. The framework is designed for easy deployment with minimal configuration.

Glioseg requires four pre-operative structural MRI scans, including T1-weighted (T1w), contrast-enhanced T1-weighted (T1CE), T2-weighted (T2w), and T2-weighted fluid-attenuated inversion recovery (FLAIR) MRI, provided in DICOM or NIfTI format (Figure 1). Preprocessing steps include (in order) reorientation and resampling to a standard space and resolution, intra-subject groupwise registration, registration to an appropriate brain atlas, and brain extraction. Five publicly available pre-trained AI models are used for multi-compartment tumor segmentation, targeting non-enhancing tumor/edema, necrosis and enhancing tumor. The specific models used are nnUNet [6], the winning submission of the BraTS 2021 challenge [7], HD-GLIO [6,8], DeepSCAN [9] and the Federated tumor segmentation (FETS) tool [10,11]. The choice of brain atlas for registration (MNI152 [12] or SRI24 [13]) is determined by the input requirements of each model. Segmentation outputs are postprocessed, including isolated voxel removal and volume-based relabeling, followed by ensembling through simultaneous truth and performance level estimation (STAPLE) [14]. Final segmentations are transformed back to native patient space to facilitate downstream analysis.

Glioseg was evaluated on an external dataset of 43 patients with adult-type diffuse glioma of varying WHO CNS5 (2021) [15] types and grades. All scans completed the pipeline without failure. Segmentations were compared in native space to semi-automatic ground truth annotations, where outputs from an external AI model were corrected by radiologists. Glioseg achieved an average whole tumor Dice score of 0.85 ± 0.09 and a multi-class Dice score of 0.77 ± 0.12 (Figure 2). The average whole tumor Dice scores of the individual models ranged from 0.78 to 0.86, while multi-class Dice scores ranged from 0.70 to 0.80. Figure 3 shows a representative case with high segmentation agreement, while Figure 4 highlights a common segmentation error seen, in which extensive non-enhancing tumor/edema is mislabeled as necrosis.

Glioseg addresses several limitations of current segmentation tools by supporting multiple scan formats, automating pre- and postprocessing, and running five models end to end in a single workflow. However, the STAPLE ensemble did not outperform the strongest individual model (BraTS ‘21), suggesting that the current ensembling strategy may not fully leverage complementary model strengths and could be hindered by lower-performing models. Future work will explore alternative label fusion methods, such as reliability-weighted ensembling or leave-one-out strategies to identify and exclude detrimental models. A persistent error was the misclassification of non-enhancing tumor/edema as necrosis. This issue is likely an inherent limitation of the individual models, which were primarily trained on whole-tumor, tumor-core, and enhancing-tumor labels without explicit differentiation between necrosis and non-enhancing tumor. Our current postprocessing only partially mitigates this problem. To improve, we will explore intensity-based or structural constraints, such as enforcing a rule that necrosis must be encapsulated by enhancing tumor. Additional ongoing work includes benchmarking against fully manual expert segmentations and established tools in the literature, stratifying performance by tumor type and size, and expanding the pipeline to include post-operative segmentation models and models that are robust to missing structural scans.

Glioseg provides a fully automated solution for clinically meaningful multi-compartment glioma segmentation. Its combination of standardized preprocessing, multi-model inference, and ensemble-based fusion in a single framework ensures reliability and user-friendliness. Ongoing development aims to further enhance performance and facilitate reproducible segmentation outcomes. With continued validation, Glioseg has the potential to become a standard tool to support glioma research.

NeuroQuantⓇ MS (NQ-MS, formerly LesionQuant) is a medical software for detecting, counting, and measuring multiple sclerosis (MS) lesion volumes seen as hyperintensities in T2w-FLAIR images, categorizing them by brain regions–(cortico)juxtacortical, periventricular, infratentorial, and deep-white-matter–based on McDonald criteria[1], while also calculating lesion burden. In 2020, Brune et al.[2] from our MS research group in Oslo found NQ-MS useful in aiding visual MS lesion detection and improving brain atrophy identification, the latter being beyond the scope of this study. More recently, our NQ-MS software was upgraded from version 3.1 to 4.12, integrating a hybrid machine learning (ML)/deep learning (DL) algorithm for enhanced segmentations[3]. This study compares the two versions using a locally collected multiple sclerosis (MS) dataset, hypothesizing that ML/DL segmentation improves lesion detection, assessed visually and through statistical analysis.

Sixty-three subjects with MS were scanned on a 1.5T Siemens Aera MRI scanner using a standard 20-channel head coil, with T2w-FLAIR and T1w-MPRAGE (for brain segmentation) and image parameters closely following NQ-MS recommendations. The dataset is part of a larger study approved by the hospital’s local ethical committee. Lesion detection volume was set to a minimum of 3 mm³ with 1 mm separation[2]. In NQ-MS version 3.1, sensitivity thresholds (range: 0–2, high to low) were applied. First the default of 1.0 (neutral), then adjusted to 1.2 (neutral-) for periventricular lesions to avoid over-detecting non-lesion structures. Visual assessments of some patients have been done, whereas all the segmentations will be reviewed by two radiologists in time for the conference. Statistical analysis of expert versus program analysis will be done using Python 3 with the Pandas module[4]. Cohen’s Kappa results will be analyzed for lesion classifications, and ICC/CCC for lesion filling percentages in NQ-MS segmentation masks. Statistical analyses of the lesion counts and volumes were conducted using R Software (version 4.3.0)[5]. The comparison of performance between the two NQ-MS versions was implemented through Wilcoxon signed-rank tests for each brain region.

Preliminary analysis of expert assessment versus NQ-MS segmentation shows a significantly larger agreement between expert and v4.12 than its predecessor. This pertains both to lesion type classification and lesion filling grade. There was also a significant preference by the expert for v4.12 segmentation. Figure 1 presents boxplots of lesion counts and volumes for the two NQ-MS versions, demonstrating statistically significant increases in v4.12 compared to v3.1, except for periventricular lesion counts. As the NQ-MS v3.1 manual suggests that a sensitivity of 1.2 may be more appropriate for detecting periventricular lesions – since a setting of 1.0 might capture non-lesion tissue structures – we reprocessed the dataset with this adjusted sensitivity. The corresponding results are shown in Figure 2.

Our study demonstrates that upgrading from NQ-MS v3.1 to v4.12, which incorporates ML/DL, results in a significantly higher sensitivity in lesion detection and volumetry as evaluated with statistical tests. NQ-MS v4.12 consistently detected more brain lesions and with larger volumes than v3.1 for the juxtacortical, deep white matter, and infratentorial anatomical regions. Using default sensitivity settings, v3.1 quantified a greater number of periventricular lesions, but not a larger volume load. However, using a lower periventricular sensitivity setting of 1.2 vs 1.0 as recommended by Cortechs, periventricular lesion counts were not statistically different for the two versions. A too high sensitivity may give false positive results though; Cortechs has likely kept their default sensitivity setting to avoid false negatives. When using the ML/DL algorithm, whether having no such settings is an advantage needs yet to be confirmed by a radiologist with MS expertise. Our preliminary expert verification and subsequent analysis confirms that the v4.12 segmentations are significantly better classified and filled than the ones of its predecessor.

Our findings highlight the enhanced performance of NQ-MS v4.12 driven by the integration of ML/DL, and underscore the significance of improving lesion quantification in MRI analysis. These advancements hold promise for improving the diagnosis and monitoring of neurological conditions. Further research and validation are necessary to explore the full potential of NQ-MS in the clinics.

Autism spectrum disorders (ASD) are a group of neurodevelopmental conditions characterised by cognitive and/or behavioural impairments. The diagnostic process is multidisciplinary and time-consuming, mainly involving neuropsychological and imaging assessments. Therefore, achieving a fast and precise diagnosis remains challenging. In this context, the use of innovative methodologies for analysing intrinsic functional magnetic resonance imaging data (fMRI), such as non-standard connectivity metrics, deconvolution approaches to mitigate hemodynamic responses, automated diagnosis and biomarker identification of ASD, among others, are promising. The purpose of this work is to evaluate and compare intrinsic fMRI data from subjects diagnosed with ASD and typical controls. To this end, five non-standard functional connectivity metrics, two deconvolution approaches, and seven machine learning classifiers for automated diagnosis were considered.

Raw data were obtained from the publicly available database ABIDE I Preprocessed (Autism Brain Imaging Data Exchange) [1]. Intrinsic fMRI timeseries (AAL atlas, 116 ROIs) [2] from 346 ASD subjects and 369 typical controls were obtained. As shown in Table 1, the following methodologies were applied: six functional connectivity metrics, namely Pearson’s correlation (PC) and five additional non-standard metrics (MULAN toolbox) [3]; two approaches for the deconvolution of the BOLD signal, namely Blind Deconvolution (BD) (MATLAB® rs-HRF toolbox) [4] and Paradigm Free Mapping (PFM) [5]; and seven machine learning classifiers to perform the automated diagnosis of ASD (Malini toolbox) [6]. For each connectivity metric, two different methods were adopted: bivariate, considering only the pairwise connections; and partial, calculating partial results to consider the influence of other ROIs. The connectivity direction was also considered: U and D indicate undirected and directed connectivity, respectively [3] (Figure 1). The following performance metrics were considered for the machine learning classifiers: accuracy, precision, recall/sensitivity, specificity, balanced accuracy and F1-score. The statistical analyses focused on accuracy included: Friedman test to compare the machine learning classifiers with Nemenyi as a post-hoc test; two-way ANOVA to evaluate statistically significant differences in the results of RLR classifier with Bonferroni as a post-hoc test.

In this work, 228 different combinations were obtained for the evaluation of the automated diagnosis of ASD using the machine learning classifiers (main results in Table 2). In terms of statistical comparison of the classifiers for accuracy, RLR and LinearSVM demonstrated significantly superior performance, while NaïveBayes and KNN exhibited significantly inferior performance (p<0.001). B computations led to higher accuracies compared to P computations. Regarding the RLR classifier, the metrics BPC, BCorrD, BH2D, BH2U, and BCorrU showed significantly higher accuracies; conversely, the metrics PCorrD-a, PCohW1-a, PCorrU-a, PCohW2-a, and PCorrU-b demonstrated significantly lower accuracies (p<0.001). In relation to the deconvolution approaches, non-deconvolved exhibited significantly superior performance when compared to BD (p=0.0036). As shown in Table 3, regarding the functional connectivity metrics, non-deconvolved demonstrated a superior performance overall, except for TE metric. BD performed better for metrics such as PC, Corr and H2. Conversely, PFM performed better for metrics of the entropy family, such as MIT and TE, as well as PC.

Improving the automated diagnosis of ASD using intrinsic fMRI data remains a major challenge due to the disorder’s complexity. In relation to the global accuracy results, the most relevant values achieved presented a performance that is similar to or better than those observed in other studies [7], in particular considering RLR and LDA classifiers. Surprisingly, the highest accuracies were obtained in the non-deconvolved group. In addition, non-standard metrics such as Corr, H2 and MIT demonstrate potential and may be a particularly useful strategy in the search for biomarkers in ASD. A considerable limitation of this work is related to the ABIDE database: although it provides a large number of subjects, which is positive for machine learning computations, it is also heterogeneous, which makes it more difficult to identify ASD biomarkers. For future work, several strategies can be explored, such as optimising parameters of the classifiers already applied, using additional classifiers, and implementing the feature selection approach before the machine learning computations.

Improving the automated diagnosis of ASD is particularly challenging; however, several strategies remain to be explored. The highest accuracies obtained in this work are similar to or higher than those observed in other studies. Corr, H2 and MIT demonstrate potential to contribute to the search for biomarkers in ASD.

Independent component analysis (ICA) has become an essential technique for denoising functional magnetic resonance imaging (fMRI) data by decomposing the signal into statistically independent components [1,2]. A critical challenge is accurately distinguishing between components representing blood-oxygen-level-dependent (BOLD) signal and non-BOLD noise components [3,4]. Recently, a new framework for component classification has been proposed for the multi-echo fMRI acquisition, which captures the signal decay across multiple echo times [5,6]. The TE-dependent analysis (TEDANA) pipeline implemented the framework, which derives selected metrics from the multi-echo fMRI data and classifies them using a custom decision tree [7]. This study builds upon TEDANA classification efforts by evaluating multiple classifiers trained on metrics from different representations of complex multi-echo fMRI data, investigating whether incorporating metrics from real and imaginary parts improves classification performance.

The classification dataset was created from manual rating of ICA components by three independent raters, using ICA component metrics derived from the magnitude, real, and imaginary parts of complex multi-echo fMRI data. Metrics included both parameters defined in TEDANA (such as kappa and rho) and custom-made parameters. After removing observations with missing values and components without clear BOLD/non-BOLD classification, the final dataset comprised 998 components: 903 non-BOLD (90.5%) and 95 BOLD (9.5%). We addressed this class imbalance by down-sampling the majority class in the training set to achieve a 1:1 ratio. Variable selection was based on non-parametric correlation analysis. Data were split into training (90%) and testing (10%) sets. Four classifiers were trained and optimized using k-fold cross-validation: Naive Bayes, Logistic Regression, Support Vector Machine (SVM), and Random Forests. Each classifier was trained twice: first using only magnitude-derived metrics, then using all metrics.

Classification using only magnitude-derived metrics yielded high test accuracies: Naive Bayes (96.0%), Logistic Regression (97.0%), SVM (92.0%), and Random Forests (94.0%). All classifiers demonstrated high precision for non-BOLD components (98.0%-100.0%) but lower precision for BOLD components (57.0%-82.0%). When trained on all metrics, test accuracies remained similar: Naive Bayes (97.0%), Logistic Regression (93.0%), SVM (89.0%), and Random Forests (91.0%). Only Naive Bayes classifier showed an improvement (+1.0%), while other classifiers showed decreases in accuracy: Logistic Regression (-4.0%), SVM (-3.0%), and Random Forests (-3.0%). Precision for non-BOLD components remained high across all classifiers when using all metrics, but precision for BOLD components decreased for most classifiers compared to using only magnitude-derived metrics.

Our findings suggest that incorporating metrics from real and imaginary parts of complex multi-echo fMRI data does not consistently improve the classification of ICA components. While Naive Bayes showed slight improvement with all metrics, other classifiers demonstrated decreased performance. The decreased performance when including all metrics may be attributed to redundant or noisy features complicating decision boundaries, consistent with low correlations between real/imaginary-derived metrics and the outcome variable (mean correlation: 0.178 for real-part, 0.077 for imaginary-part). This study analysed a single train-test split, with variability observed across different training subsets. The high class imbalance remains a challenge despite the down-sampling approach. Future work should include aggregated analyses across multiple train-test splits and feature selection methods to identify the most informative combination of metrics. Overall, magnitude-derived metrics remain most valuable for BOLD vs. non-BOLD classification in multi-echo fMRI data.

This study demonstrates that metrics derived from the magnitude representation of complex multi-echo fMRI data provide robust features for classifying ICA components as BOLD or non-BOLD. The overall classification accuracy was 92% or higher. The inclusion of metrics from real and imaginary parts of the data did not yield consistent improvements in classification performance across different algorithms. The work on this publication was supported by the Czech Science Foundation (GAČR), project No. GA23-06957S.

At the latest with the publication of the nnU-Net framework [1] and the availability of large cohort studies including magnetic resonance (MR) imaging (e.g., UK Biobank [2] or the German National Cohort, NAKO [3]), large-scale processing of MR images, for example, segmentation of various different target structures (e.g., organs, bones, muscles, or adipose tissue) is widely performed eventually leading to versatile multiclass segmentation models, e.g., TotalSegmentator [4] and related systems [5–7]. The training data sets of these models often show dependencies, i.e., different iterative generations of MR segmentation models are based on output segmentations of former models, eventually leading into a vicious cycle. This could in part be a direct consequence of the time-consuming manual annotation process, usually requiring special training. Recently, Willem et al. showcased the different biases that can occur along the general machine learning (ML) pipeline [8]. Segmentation performance of (parts of) the human body is directly influenced by its natural variability, besides the primary category of sex along the continuous scales of age and body size (e.g., using body mass index [BMI]). Additionally, important categorical data such as race/ethnicity is seldom collected in a standardized way or underrepresented. It can be assumed that poorly curated training data sets lead to less generalizable segmentation models. Bias introduced by using such data sets is referred to as cohort bias. Aim of this study was to systematically investigate the influence of cohort bias in the training data set of segmentation models using the example of adipose tissue (AT) segmentation.

3D VIBE Dixon MRI from NAKO (protocoll details to be found in [3]) were used to create 18 distinct training data sets (3 age groups [20–39, 40–59, >60 years], 3 BMI groups [18.5–24.9, 25–29.9, >30 kg/m²] for women/men, respectively) with n=30 each. These sets were used to train 18 distinct nnU-Net models (1 fold, nnU-Net ResEnc M [9], 3d_fullres configuration) [10]. An additional randomly drawn test set from NAKO (n=1000) was used to estimate model performance on the general population. Each model was tested on the 17 remaining biased training data sets and on the test set in terms of segmentation performance (DSC for 2 classes; subcutaneous AT [SAT], visceral AT [VAT]) and AT volumes representing real-world outcomes.

Fig. 1 shows the on-average top/bottom 3 models ranked by DSC. On average, the best performing model was trained on data from women between 40 to 59 years with BMI >30 kg/m². By contrast, using data from men <60 years with BMI between 18.5–24.9 kg/m² led to the worst performing models. Fig. 2 summarizes intersectional testing of biased models. On the test set from the general population, mean DSC was 0.977 for SAT and 0.964 for VAT. Volumetric errors range from 0.2–2% (mean 0.7%) for SAT and 0.5–6% (mean 1.8%) for VAT. Best performing model for both classes was trained on overweight women (20–39 years).

The composition of the training set affects the performance of the segmentation models. In the present study, n=30 image datasets were used per training dataset. This number has led to very reliable models in the past [10]. It can therefore be assumed that the influence of unbalanced training sets is even stronger in a scenario with limited availability of image datasets, especially in anatomically complex structures, e.g., visceral adipose tissue.

Using the example of AT segmentation, it was shown that the composition of the training set according to relevant parameters (in this specific case sex, age and BMI) plays a major role in the development of generalizable segmentation models.

MRI images allow identification of spinal nerve rootlets and intervertebral discs (IVDs), which can be used to determine spinal and vertebral levels, respectively [1, 2]. Spinal rootlets are relevant for functional MRI group analyses [3, 4] and neuromodulation therapy [5]; however, their applicability is impacted by considerable variability in spinal and vertebral levels across individuals [2, 6]. A method for automatic rootlet segmentation was proposed [1], but it is limited to a single contrast and dorsal rootlets only. In this work, we developed a model to segment dorsal and ventral C2–T1 rootlets from MP2RAGE and T2w MRI data. Then the model was used to explore the correspondence between spinal and vertebral levels.

Three datasets of healthy adults covering the cervical spinal cord were used for model training: an unpublished dataset of 7T MP2RAGE (3 aligned contrasts: UNIT1, INV1, INV2) 0.7×0.7×0.7 mm data (15 training and 4 testing participants resulting in 45 training and 12 testing images) and two open-access datasets of 3T T2w 0.6×0.6×0.6 to 0.8×0.8×0.8 mm images (spine-generic [7]: 21 training and 3 testing images; and OpenNeuro ds004507 [8]: 10 training and 2 testing images). Ground-truth (GT) labels of dorsal and ventral C2-T1 rootlets were generated semi-automatically for all training and testing images using an existing segmentation method [1] (which segmented only dorsal C2-C8 rootlets), followed by manual correction and annotation of ventral rootlets using FSLeyes viewer. A multi-contrast segmentation model was developed with the nnUNet framework [9] using 76 training images, trained with 2000 epochs, and 17 testing images with GT masks to compute Dice. To analyse the correspondence between spinal and vertebral levels, we used all the training and testing data and an additional 82 images [7] automatically segmented by the proposed model. Figure 1 shows the automatic analysis pipeline utilising Spinal Cord Toolbox (SCT) v6.4 [10]. For each image, the spinal cord was segmented [11, 12], IVDs and the pontomedullary junction (PMJ) were identified [13, 14], and nerve rootlets were segmented using the proposed model. Then, IVDs were used to estimate the vertebral levels and rootlets were used to obtain the spinal levels [1]. Lastly, the distance between the PMJ and spinal and vertebral level midpoints was measured along the cord centerline. To account for participants' demographics, we normalised the distances by the participant’s height and multiplied them by the median height (to preserve millimetre units).

Figure 2 shows illustrative segmentation by the model for one T2w and one MP2RAGE test image. The model achieved a Dice score of 0.65 ± 0.10 (mean ± standard deviation [SD] across levels and 17 testing images). Figure 3 presents the Dice across levels and contrasts on the testing images. Figure 4a depicts the correspondence between spinal and vertebral levels measured as the distance from the PMJ. Figure 4b shows the Bland-Altman analysis. The bias term for vertebral level C2 (VLC2) and spinal level C3 (SLC3) is zero, suggesting that these levels correspond exactly.

A multi-contrast model showed a robust segmentation performance across four contrasts (T2w, INV1, INV2, and UNIT1) and different levels (C2-T1) for both ventral and dorsal rootlets. The analysis of spinal and vertebral level correspondence confirms the previous MRI report [2] and anatomical textbooks [15] that the spinal levels are shifted relative to the vertebral levels and that this shift changes with lower levels. The more caudal the level, the greater the shift (Figure 4a and bias term in Figure 4b across levels), coupled with higher inter-subject variability (more scatter clusters in Figure 4b across levels). In this study, we used the participants' height to account for potential biological differences among individuals. Ideally, the spinal cord or spine length would be considered, but MRI data typically does not cover the entire spine.

We developed a deep-learning model for spinal rootlets segmentation across different MRI contrasts and studied the correspondence between spinal and vertebral levels in an adult population. Our results confirm that spinal levels are shifted relative to vertebral levels, and that this shift increases at more caudal levels, accompanied by greater inter-subject variability. Future work will focus on validation using data from participants with spinal pathologies.

For patients with unresectable pancreatic tumors, stereotactic magnetic resonance guided adaptive radiation therapy (SMART) increases survival by delivering a high biologically effective dose in few fractions, although many patients still experience early metastatic or local recurrence [1, 2]. To address this challenge, radiomics leverages mathematical analysis to extract and quantify textural information from medical images, thereby providing additional insights to predict such outcomes. Meanwhile, delta-radiomics assesses the variation in features captured at treatment times and has demonstrated potential in predicting radiotherapy outcomes when combined with machine learning [3]. However, in the case of pancreatic cancer, these studies lack both internal and external validation, which we aim to address here by rigorously showing the performance of such models.

We included all patients treated with SMART for unresectable pancreatic ductal adenocarcinoma from Montpellier Cancer Institute (France) and from University Hospital La Milagrosa, Madrid (Spain). For all these patients, 0.35T T1-w TrueFISP simulation (simu) and fraction (Fn) MRI were acquired. Radiomics features were extracted following IBSI guidelines [4] on the manually annotated gross tumor volume (GTV) ROI. Delta-radiomics features were calculated between different Fn and simu, and reliable features, as identified in a previous study [5], were analyzed. Various feature selection algorithms and predictive machine learning models were evaluated to predict local recurrence at 1 year (LR1y) and metastatic recurrence at 9 months (MR9m) using radiomics or delta-radiomics data from a single time-point. Prediction models were trained on 70% of the Montpellier dataset after feature selection and first evaluated with an internal validation cohort of 25 patients. A subsequent evaluation was performed using an external validation cohort of 37 patients from Madrid, where radiomics features were harmonized using the ComBat method [6]. Both evaluations were performed using bootstrapping (n=200) to compute the 95% confidence intervals for the area under the curve (AUC) and the Brier score.

85 patients were recruited from Montpellier Cancer Institute (France) and 37 from University Hospital La Milagrosa, Madrid (Spain). In the first cohort, 14 presented LR1y and 44 MR9m status while they were 5 and 15 respectively in the second cohort. In the internal validation group, the Random Forest (RF)+RF model using F3 radiomics data achieved the highest AUC of 0.96 (95% CI: 0.87-1.00) for predicting LR1y, with a Brier score of 0.08 (95% CI: 0.03-0.14), 84% sensitivity, and 100% specificity. For MR9m, the RF+ADABOOST model with F1/F3 delta-radiomics data yielded an AUC of 0.83 (95% CI: 0.66-0.97), a Brier score of 0.22 (95% CI: 0.20-0.24), 91% sensitivity, and 75% specificity. In the external validation cohort, these models demonstrated an AUC of 0.71 (95% CI: 0.54-0.84) for LR1y and 0.51 (95% CI: 0.28-0.75) for MR9m, with Brier scores of 0.19 (95% CI: 0.11-0.27) and 0.25 (95% CI: 0.22-0.29), respectively. Sensitivity and specificity were 40% and 78% for LR1y, and 40% and 68% for MR9m. In the external validation cohort, the ANOVA K BEST+PSVM model using Simu radiomics data achieved the highest AUC of 0.80 (95% CI: 0.50-0.99) for LR1y, with a Brier score of 0.25 (95% CI: 0.23-0.27), 80% sensitivity, and 56% specificity. For MR9m, the ANOVA K BEST+RF model using F1/F3 delta-radiomics data achieved an AUC of 0.85 (95% CI: 0.71-0.96), a Brier score of 0.23 (95% CI: 0.22-0.24), 73% sensitivity, and 82% specificity.

This study demonstrates the potential of delta-radiomics data combined with machine learning models to predict radiotherapy outcomes for patients with pancreatic tumors. Rigorous internal and external validation was crucial for assessing model reliability and generalizability. While LR1y predictions showed strong performance in internal validation, they significantly declined in the external cohort due to the low incidence of local recurrence, which highlighted the need for recalibration [7]. In contrast, MR9m predictions, although not as strong as LR1y, maintained robust performance with F1/F3 delta-radiomics models. This is particularly encouraging, as metastatic recurrence was the primary cause of mortality in the Montpellier dataset (23 out of 34). Finally, the application of the ComBat method for harmonizing radiomics features was limited by small subgroup sample sizes (n<20), preventing its full utilization.

This study highlights the potential of delta-radiomics combined with machine learning to predict radiotherapy outcomes in unresectable pancreatic tumors, demonstrating the importance of rigorous validation. Future work should prioritize the inclusion of additional cases and the enhancement of model calibration. Incorporating clinical and dosimetric features could further augment predictive power and address the challenges observed in the external validation.

Breast cancer is the most common cancer among women worldwide, reaching almost 2.3 million new cases in 2022 [1]. Early detection is critical, and compared to mammography breast magnetic resonance imaging (MRI) shows higher sensitivity, particularly for identifying invasive cancers [2]. Recent advances have focused on ultrafast contrast-enhanced (UF-DCE) MRI protocols, which offer dynamic imaging capturing early enhancement kinetics, enabling potentially better differentiation between benign and malignant lesions [3]. Growing interest has been reported in multimodal AI systems that integrate imaging and non-imaging data to better capture the complexity of patient profiles [4–6], to enhance diagnostic accuracy. Thus automatic multimodal models have been developed to classify breast lesions taking advantage of UF-DCE MRI images, patient data extracted from written reports, and geometric features computed from manual segmentations of lesions [7]. One caveat of these models is the requirement for manual time-consuming segmentation to compute geometric features. Automatic breast lesion segmentation in conventional DCE-MRI, remains challenging due to the low signal-to-noise ratio and significant inter-patient enhancement variability [8]. In this work, we investigate the use of automatic segmentation models in UF-DCE to automatically extract these geometric features.

As shown in Figure 1, the main goal is to remove the need for human intervention in lesion segmentation to derive geometric features, exploited in our multimodal classification approach [7]. To achieve this, we used a custom U-Net model with an automatic extraction processing of geometric features. While the lesion position is still required to crop around a lesion (volume of interest) due to the absence of a detection model, this step is significantly less time-consuming than full manual segmentation. Automatic lesion segmentation: a U-Net model [9] was trained using the same images as the one used for the classification model, with a key difference in the cropping strategy. Instead of cropping only around lesions, random regions were sampled within each whole breast, excluding background areas without tissue, increasing the amount of training data for the segmentation model. Lesion position being important for the classification model, particularly in identifying lymph nodes [7], we performed a Global Position Embedding (GPE) by encoding the crop location relative to the whole breast as a three-channel 3D image (one for each dimension) and concatenated it with the 3D last phase UF-DCE subtracted images. Geometrical features extraction: for each lesion, the U-Net produced a coarse lesion segmentation map that was binarized (background = 0, suspected lesion = 1). Then, we used the SimpleITK library [10] to refine the segmentations by keeping only the region closest to the lesion location and whose size was greater than 50 voxels, as depicted in Figure 2. Using the kept region, geometric features were extracted including bounding box (position and size), volume, center of gravity, ellipsoid axes lengths, elongation, and flatness [7]. If no valid region was found, geometric features could not be extracted and the classification model replaced them with the mean value of each feature.

For all experiments, we used a stratified 5-fold cross-validation on a dataset including 987 manually segmented lesions classified as malignant, benign or lymph nodes [7]. On the segmentation task exemplified in Figure 3, when only considering crops centered on lesions, our segmentation model achieved a DICE coefficient of 60.9±1.2%. On the classification task, we investigated the impact of the automatic features extraction during the training or test stages. We compared two variants, “auto” and “hybrid”, against the reference model [7] which used features from manual segmentations (Figure 4). The auto model used a classifier that was both trained and evaluated on the automatic features, while the hybrid model was trained on the original features but evaluated on the automatic features. Resulting sensitivity, specificity and Area Under the ROC curves (AUC) are reported in Figure 4.

Both methods relaying on automatic segmentation underperformed compared to the baseline reference using manual segmentation, which was also reported in the literature [11] . However, the hybrid approach offered a promising compromise between segmentation effort and accuracy. Indeed, this method achieved an AUC only 2.8% lower than the reference, while eliminating the need for manual segmentation at inference time. Additionally, training the classifier on manual segmentations yielded a 3.6% higher AUC than training directly on automatically extracted features.

The proposed hybrid method can be used to avoid manually segmenting lesions during clinical practice at a small accuracy cost. Further work should be done to improve the segmentation model and therefore the extracted geometric features.

Diffusion MRI enables the estimation of tissue microstructural parameters performing acquisitions with varied acquisition settings. However, the high dimensionality of the acquisition parameter space can result in long acquisition times. Therefore, selecting the most informative subset of measurements to characterize tissue microstructure remains crucial. Recent work [1] proposed a physics informed machine learning approach that couples a Concrete autoencoder (CA) [2] with a signal forward equation to automatically subselect the most informative dMRI measurements. This approach has shown high accuracy of parameter estimates compared to subsampling based on random, uniform, Cramer-Rao Lower Bound (CRLB) [3] (non-physics informed machine learning strategies). In this work we further assess uncertainty and degeneracy of biophysical parameter estimation under measurement subset selection by different selection methods using µGUIDE [4], a Bayesian inference framework that estimates the posterior distributions of microstructural parameters using deep learning.

Synthetic data: The MUDI dataset [5] was used as basis for the training, it consists on a 5D Diffusion-T₁-T₂* protocol acquired on a 3T scanner (80 mT/m) using 1344 unique settings, varying b-value ([0–3000] s/mm²), gradient orientation, inversion time (TI: 20–7322.7 ms) and delay time (TD: [0, 25, 50] ms). The repetition time (TR) was 7500 ms, and the b-tensor anisotropy (bΔ) was 1. To generate synthetic data for training and evaluation, signals were simulated using the following signal equation: S = S₀·exp (-b∶ D) · |1 - 2 ·exp⁡(-TI / T₁) + exp(-TR / T₁)| · exp(-TD / T₂*) And b:D: b∶ D = (1/3) · b · b_Δ · [D_∥ - D_⊥] - (1/3) · b · [D_∥ + 2 · D_⊥] - b · b_Δ · (g(Θ_g,Φ_g) · g(θ,φ))² · [D_∥ - D_⊥] Where b is the b-value; bΔ the anisotropy factor of the diffusion encoding; D∥ D⊥ parallel and perpendicular diffusivities; g(Θg,Φg) the direction of the diffusion gradient in spherical coordinates; g(θ,ϕ) the principal direction of the tissue structure; TI, TR and TD inversion time, repetition time and echo delay time, T1 and T2∗ the longitudinal and effective transverse relaxation times. We used the following ranges to simulate the data: S_0∈ [0.5,5.0] T_1∈ [100,5000] ms T_2^*∈ [0.01,2000] ms D_∥∈ [0.01,3.2] μm2/ms D⊥=k⋅D∥D_⊥ = k⋅ D_∥, with k ∈ [0.01,1] θ∈ [0,π], ϕ∈ [0,2π] Synthetic noise was added at an SNR of 30 for the lowest b-value and longest TI/TD. Measurement subset selection and parameter estimation: Figure 1 illustrates the proposed workflow. Firstly, measurements were sub-selected using uniform, random, CRLB and physics-informed machine learning selection (CL+eq) (Figure 1 A1); then, the selected measurements were fed to uGUIDE for training and inference of the posterior distribution of microstructural parameters (Figure 1 A2). The selected measurements were fed to uGUIDE (Figure 1 B1). The number of selected measurements were N = 500, 100, 50. Finally, the maximum a posteriori, degeneracy and uncertainty were computed and compared to the full dataset.

Results and Discussion: Figure 2 presents examples of posterior distributions for the estimated parameters. In purple, the posterior distributions obtained using all the signal measurements, it can be seen the full width at half maximum (FWHM) of the distribution is narrow. The posteriors using a decreased number of measurements are shown for all 4 sub-selection methods. As the number of measurements decreases, the posterior distributions remain centered around the ground truth, with a moderate increase in FWHM. This indicates a slight increase in uncertainty, yet the estimates remain stable and well constrained. Figure 3 presents the Maximum a Posteriori (MAP) estimates for decreasing sample sizes (N=500, 100, 50). The results show that CL+eq consistently yields more accurate MAP estimates with lower uncertainty, particularly at lower sampling rates, aligning with findings of higher accuracy in previous work [1]. Figure 4 presents the uncertainty for the different sub-selection methods. Again, the CL+eq method shows the most consistent dispersion of the most probable samples, indicating robust performance across varying sample sizes. In contrast, Uniform and Random selection strategies show greater uncertainty and variability in the estimates. Although CRLB optimization is grounded in analytical minimization of variance, it is still outperformed by the CL+eq approach. An exception can be noted for T2∗ , where CL+eq shows reduced performance. This is attributed to the fact that CL+eq was trained on a representative parameter distribution in human data for measurement subset selection, whereas the evaluation here was performed on a uniform distribution including long T2*, and the delay times (TD) included in the acquisition are short.

The results reinforce the benefit of physics-informed selection strategies and highlight the value of µGUIDE in assessing the performance of reduced measurement sets.

Multiple sclerosis (MS) is characterized by white matter lesions (WMLs) that exhibit heterogeneous features, reflecting variability across different stages of demyelination, neuroinflammation, and neurodegeneration. Quantitative and semi-quantitative magnetic resonance imaging (qMRI) can help to detect these underlying pathological alterations in vivo. This study aims to identify different lesion profiles based on qMRI features using an unsupervised machine learning approach.

We included 91 MS patients (76% females; mean age 41.7 ± 13.0 years) who underwent a brain 3T MRI on Philips Elition S scanner at the University Hospital of Verona, Italy. WMLs were manually segmented on T1-weighted (T1w) and T2-weighted (T2w) FLAIR images. Mean qMRI values were extracted from the cores of WMLs with a core volume > 9 mm3. T1/T2 ratio (T1T2r) was computed from calibrated T1w (T1c) and T2w (T2c) images. Intensity calibration was performed using non-brain tissues (eye and temporalis muscle) [1]. Magnetization transfer (MT) ratio (MTR) was calculated from two 3D GRE acquisitions, with and without MT pulse [2]; MT saturation (MTS) was estimated by integrating an additional T1w image to reduce T1 relaxation time and B1 inhomogeneity effects [3]. T2 relaxation time-based metrics, Myelin, Intra/Extracellular and Free Water Fraction (MWF, IEWF, FWF), Total Water Content (TWC), and T2 of IE water (T2IE), were derived from a multi-echo GRASE sequence using the T2SPARC method [4] for multi-exponential fitting using the toolbox by Canales-Rodríguez (v0.3) [5]. Susceptibility-based metrics included χ-separation and quantitative susceptibility mapping (QSM). χ-separation exploits R2 and R2* (estimated through monoexponential fitting from GRASE and ME-GRE, respectively [6]) to estimate negative χneg (myelin) and positive χpos (iron) sources [7]. QSM was also obtained from the ME-GRE using the QSMxT software toolbox [8]. MRI acquisition parameters are summarized in Tab.1. QSM images were used to identify Paramagnetic Rim Positive and Negative lesions (PRL+/−): PRL+ lesions show a hyperintense rim at the border from iron-laden microglia, while PRL− lesions do not [9]. WMLs qMRI features were scaled using the mean and standard deviation of the normal appearing white matter (NAWM) calculated at a subject level. Principal Component Analysis was applied to the z-scored features for dimensionality reduction, followed by hierarchical clustering using the FactoMineR package (v2.11) [10]. The number of principal components (PCs) to retain was assessed with Horn's Parallel Analysis [11, 12]. The optimal cluster (CL) number was determined using the majority rule among various indices calculated by the NbClust package (v3.0.1) [13]. Differences in qMRI values between WMLs in each CL and NAWM were assessed using Kruskal-Wallis [14] and Dunn’s post-hoc test [15, 16] with Bonferroni correction for multiple comparisons. The effect size (res) was used to evaluate the magnitude of those differences. After clustering, patients were stratified by age quantiles (Q1 to Q4) to examine the distribution of WMLs among CLs. Clustering and statistical analysis were performed in R (v4.4.2) [17].

A total of 720 WMLs were analysed. Four PCs, explaining over 80% of the total variance, were selected. Hierarchical clustering was performed using Euclidean distance and the Ward method and 3 CLs were identified, with an average Silhouette of 0.26. CL1 collected lesion cores with the most altered values compared to NAWM in almost all the MRI metrics (T1T2r, R2, TWC, T2IE, MTS, MTR, FWF, R2*), which similarly contributed to PC1 (R2>0.7, p<0.001) [Fig.1]. CL2 was characterized by the highest values of χpos (res=0.42, p<0.001) and QSM (res=0.45, p<0.001) [Fig.1-2-3], which are the features most correlated to PC2 (R2=0.95 and R2=0.84, respectively, p<0.05). CL3 had less altered WMLs features compared to NAWM (p<0.001). PRL+ had a higher percentage of lesions in CL1 and CL2 (QSM: CL1 56% CL2 35%) with respect to PRL- (CL1 23%; CL2 24%). CL2 lesions number decreased with age (Q1 47.0%, Q4 7.5%), while CL1 and CL3 lesions number increased (Q1 21.9% vs Q4 36.2%; Q1 31.1% vs Q4 56.3%, respectively).

Clustering analysis identified 3 distinct WMLs profiles: CL1, characterized by qMRI features indicative of severe tissue disruption more frequently observed in the core of PRL+ lesions; CL2, marked by elevated χpos and QSM values suggestive of iron accumulation in the lesion core consistent with active lesions; CL3, showing milder qMRI alterations relative to NAWM indicative of inactive/remyelinated lesions. The higher prevalence of WMLs in CL2 in younger individuals supports the association of this CL with active lesions, as they seem to decrease with age.

Unsupervised clustering based solely on multiparametric qMRI identified distinct profiles of WMLs, reflecting different patterns of tissue alteration and highlighting qMRI's utility in capturing lesion heterogeneity in MS.

Magnetic Resonance Imaging (MRI) is a valuable method for detailed body composition analysis, providing detailed quantification of fat and muscle distribution, which is essential for assessing metabolic health [1]. AMRA’s Body Composition Profile (BCP) analysis by segmentation provides clinically relevant data on body fat and muscle composition. It is based on following protocol: six rapid 3D-GRE DIXON imaging stacks, three 14-second LAVA FLEX breath-hold stacks from the neck to pelvis, and three 30-second LAVA FLEX free-breathing stacks from pelvis to knee. This standard protocol, however, relies on the built-in T/R body coil, consequently disabling parallel imaging techniques, thus constraining spatial resolution and/or shorter acquisition times. To address these limitations, our study investigates the development of an advanced BCP protocol incorporating the AIR™ Coil and leveraging the latest MRI advancements, including deep learning-based reconstructions. Validating faster and optimized protocols could streamline MRI-based body composition assessments enhancing efficiency, patient comfort, and clinical utility.

Five healthy volunteers (3 males and 2 females) were scanned using standard BCP protocol with traditional T/R [BODY Coil] as well as three alternative protocols using two 30-Ch AIRTM Coils, along with a built-in posterior array AIR coil: (1) BCP protocol with parallel imaging activation thanks to the use of multi-element coils [AIR Coils] (2) BCP protocol reconstructed with a prototype of AIRTM Recon DL algorithm for DIXON aiming to denoise, remove truncation artifacts in all directions while simultaneously interpolating images [ARDL] and (3) BCP protocol acquired with a new deep learning-based reconstruction algorithm designed for highly undersampled data [DLS]. This algorithm enables increased in-plane spatial resolution keeping similar scan duration. Protocol details can be found in Table 1. For each protocol, DICOM images were transferred to AMRA (Linköping, Sweden), on which body composition analyses were conducted with AMRA® Researcher. Four body composition parameters were obtained for each volunteer: Abdominal Subcutaneous Adipose Tissue volume - ASAT (l), Visceral Adipose Tissue volume - VAT (l), Mean Anterior Thigh Muscle Fat Infiltration - MFI (%), and Total Thigh Fat-free Muscle volume - FFMV (l). To determine whether the alternative accelerated protocols provide results comparable to the standard protocol, the agreement between the standard BCP and the alternative protocols was assessed by calculating the mean differences and the limits of agreement (±1.96*SD). The standard BCP with BODY Coil was considered the reference method, while the BCP with AIR Coils, BCP with ARDL, and BCP with DLS were considered alternative methods. Details can be found in Figure 1.

Figures 2 and 3 illustrate typical examples of image segmentation based on the protocols described above. ASAT: x̄ BODY Coil: 3.85l; AIR Coil, Bias: 0.03, Limits of agreement [-0.10, 0.16]; ARDL, Bias: -0.05, Limits of agreement [-0.21, 0.11]; DLS, Bias: -0.01, Limits of agreement [-0.14, 0.11]; VAT: x̄ BODY Coil: 1.51l; AIR Coil, Bias: 0.01, Limits of agreement [-0.15, 0.17]; ARDL, Bias: 0.02, Limits of agreement [-0.14, 0.17]; DLS, Bias: 0.02, Limits of agreement [-0.12, 0.15]; MFI: x̄ BODY Coil: 4.10%; AIR Coils, Bias: -0.01, Limits of agreement [-0.18, 0.16]; ARDL, Bias: -0.49, Limits of agreement [-0.75, -0.23]; DLS, Bias: -1.16, Limits of agreement [-1.33, -0.99]; FFMV: x̄ BODY Coil: 11.79l; AIR Coils: Bias: -0.05, Limits of agreement [-0.18, 0.09]; ARDL: Bias: -0.13, Limits of agreement [-0.22, -0.03]; DLS: Bias: -0.03, Limits of agreement [-0.19, 0.12].

Analysis of the results indicates that all alternative protocols show a slight bias relative to the standard BCP protocol using T/R BODY coil. BCP with Air Coils and parallel imaging activation demonstrated the closest agreement across ASAT, VAT, MFI and FFMV measurements, with minimal bias and narrow limits of agreement, suggesting it may be a reliable alternative. The BCP with DLS aligned well with the standard BCP but showed slight variability, particularly in MFI measurements – possibly due to the higher resolution. ARDL algorithm, while generally consistent, tended to underestimate values for FFMV. Overall, the variability between protocols was small and comparable to intra-scanner test-retest variability of the standard BCP protocol, except for MFI where the variability was comparable to inter-scanner variability [2].

Applying AIR Coils shows the best agreement with the standard BCP, offering a faster, reliable option for body composition profiling. The BCP with DLS or ARDL provide a much faster viable alternative with slight limitations at this moment. These findings suggest AIR Coils with DL reconstruction could replace the BCP standard, improving efficiency, comfort, possibly extendable to whole body MRI applications, by dramatically reducing scan time.

Coronary magnetic resonance angiography (MRA) is a technique used to obtain morphological information about coronary arteries by collecting data during their stationary periods. The stationary period of coronary arteries is visually determined by an operator using cine images; however, this method is time-consuming and operator-dependent. Therefore, attempts have been made to automate the detection of stationary periods. Automatic detection of the stationary period of coronary arteries consists of two steps. In Step 1, the position of the coronary arteries is detected, and a motion curve is generated based on the positional changes across frames. In Step 2, the stationary period is obtained from the motion curve. In a previous study proposing a method using deep learning, the stationary period was determined using a threshold method after the position of the coronary artery was detected using deep learning [1, 2]． We previously proposed an automatic technique that uses deep learning for position detection and stationary period determination [3]. It uses four-chamber cine images to detect the positions of the coronary arteries using a single-shot multibox detector (SSD), a type of convolutional neural network, and then performs image segmentation using U-Net on two-dimensional regions composed of the motion curves of the left and right coronary arteries. In this study, we investigated whether it is possible to automatically determine the stationary period using a time-series prediction deep-learning model for cardiac cine images, which are essentially time-series data.

Cine images of 31 to 91 phases were acquired using 1.5-T and 3.0-T magnetic resonance imaging systems, and the positions of the coronary arteries were detected using SSD for a total of 1,031 cine images. Next, the motion curve of the coronary artery was calculated from the change in the position of the coronary artery between frames and used as the training and test data. The deep-learning neural networks used to detect the stationary period include long short-term memory (LSTM), gated recurrent unit (GRU), and transformer, which are representative models for time-series data. The evaluation of stationary period detection was conducted using recall, specificity, and their geometric mean (G-Mean). The timing of the start and end of the stationary period was represented by the phase number of the cine image, and the detection error (in phases) was obtained by calculating the mean squared error between the detected stationary period and the stationary period (label) visually evaluated by an experienced operator. The detection error (in phases) was calculated and evaluated at eight points, four points each on the left and right, namely, the start and end of the systole and diastole of the coronary artery. The computer used had 32GB of memory, Core i7 13620H CPU (Intel Corporation), an NVIDIA RTX-4070 GPU, an ADAM optimizer, and a learning rate of 0.001. Training was performed until the minimum loss was achieved, and the number of epochs was 8600 for GRU, 10,550 for LSTM, and 20,000 for transformer.

The results for GRU, LSTM, and transformer were 0.93, 0.93, and 1.00 for recall, 0.87, 1.00, and 0.99 for specificity, and 0.90, 0.96, and 0.99 for G-Mean, respectively (Table 1). The detection error in mean squared phase errors was 2.17 for GRU, 0.99 for LSTM, and 0.70 for transformer. Except for specificity, which was almost the same as LSTM, transformer showed the best results in all other indices (Fig 1-3, Table 1).

Between GRU and LSTM, the latter showed better results. GRU has a simpler gate structure than LSTM does; therefore, the computational load is lighter. However, LSTM is thought to be superior in terms of long-term memory retention. This might have been the reason for LSTM performing better than GRU in terms of memory of the various coronary artery movements associated with heartbeat, including individual differences. However, transformer showed the best results in all other indices except for specificity, compared with GRU and LSTM. Unlike GRU and LSTM, transformer does not use a sequential process but uses an “attention mechanism” that focuses on all time-series data at once. Therefore, transformer can efficiently focus on the positions of the stationary period in the entire cardiac phase, which is thought to enable a more stable and accurate detection of stationary periods. However, because the results of this study were based on a relatively small amount of data, further studies with larger amounts of data are required.

In this study, the deep-learning time-series prediction models LSTM, GRU, and transformer, which handle time-series data, were demonstrated to be capable of estimating coronary artery stationary periods. In particular, transformer demonstrated high accuracy in detecting the stationary period of the coronary arteries and is expected to be useful in improving image quality and reproducibility of coronary MRA.

Intramuscular fat fraction (FF), assessed with quantitative MRI (qMRI), has emerged as one ofthe few responsive outcome measures in CMT1A patients. The main limitation for its use in futuretrials is the time required for the manual segmentation of individual muscles. This study aimed atassessing the accuracy and responsiveness of a fully automatic Artificial Intelligence-based (AI)segmentation pipeline to evaluate disease progression in a cohort of CMT1A patients.

Twenty CMT1A patients were included in this observational prospective longitudinal study. FF was measured twice a year using qMRI in the lower limbs. CMT1A patients were positioned supine while the leg and thigh of the non-dominant limb were imaged using a combination of flexible coils on the top and spine coils integrated into the scanner bed on the bottom. After a localizer, sets of images were recorded in the axial plane at 1.5T (MAGNETOM Avanto, Siemens Healthineers, Erlangen, Germany) during approximately 45 minutes. Anatomical and quantitative imaging sequences were used in order to compute Fat Fraction (FF) through the generation of quantitative FF maps. The same protocol was conducted at 12 months for the follow-up evaluation (Figure 1). Muscle segmentation was performed fully automatically using a trained convolutional neural network with or without a human quality check(QC). The corresponding results were compared with the Dice similarity coefficient (DSC) with those obtained from a fully manual (FM) segmentation. In each case, FF progression and its standardized Response Mean (SRM) was computed in individual muscles over the single central slice and a 3D muscle volume at thigh and leg levels to define the most sensitive region of interest.

The technical performances and limitations of manual segmentation were assessed through an initial analysis of the between- and within-operator variations and so in a sample of ten CMT1A thigh level images from our cohort. The corresponding DSC values between the two operators were very high (0.92 ± 0.02) (Figure 2). AI-based segmentation showed excellent DSC values (>0.90). A significant global FF progression was observed at thigh (+0.71 ± 1.28%; p=0.016) and leg levels (+1.73 ± 2.88%, p=0.007), similarly to what was computed from the FM technique (p=0.363 and p=0.634). FF progression of each individual muscle was comparable when computed either from the central slice or the 3D volume. In order to identify the most sensitive region of interest (ROI), FF progression rates together with the SRM values were assessed using AI-based segmentations methods in a single central slice and within the whole 3D volume. The optimal SRM value (0.70) was actually obtained for the global leg 3D volume analysis using the AI+QC segmentation method. The time necessary for the fully automatic segmentation process using AI with a QC was10 hours for the entire dataset as compared to 90 hours for the FM.

We initially appraised the performance of the CNN-based automatic segmentation process considering the manual segmentation technique as the ground truth. The accuracy of the AI-based segmentation was excellent as illustrated by the >0.90 DSC value obtained for all analyzed muscles. This excellent DSC value was similar to what we previously described and slightly better than the values reported in the literature with other CNN. These values should be definitely analyzed in view of the limitations of the manual segmentation process. The absence of any significant difference in our study between the FM and the AI-based segmentation techniques regarding the FF progression rate over one year supports that qMRI coupled with AI-based segmentation techniques could be used in future longitudinal studies in CMT1A patients. As such, slowly progressive disease such as CMT1A could be followed over short periods of time on the basis of FF changes computed using qMRI and AI-based segmentation. Given the slow FF progression at thigh level and the large heterogeneity between muscles and individuals, FF should be quantified from a 3D volume at the leg level for longitudinal analyses.

In conclusion, AI-based segmentation technique applied to neuromuscular qMRI is ready to be used in longitudinal studies to assess disease severity and progression in CMT1A patients. Of interest, such an approach makes it possible a 3D volume and individual muscle analysis in the lower limbs. The addition of a QC step is still recommended as it might further improve the FF responsiveness in follow-up studies with higher SRM.

Gliomas are a common primary brain tumor, originating from glial cells, and are associated with poor prognosis, particularly in high-grade forms like glioblastoma [1]. Magnetic Resonance Imaging (MRI) is crucial for diagnosis, offering detailed structural images without ionizing radiation. However, manual segmentation of MRI scans is time-consuming and prone to inter-observer variability, affecting diagnostic accuracy [2]. Deep learning models, such as YOLO, have been applied for automated tumor detection in medical imaging, with YOLOv11 [3] demonstrating improved speed and accuracy. Despite these advancements, many models analyze MRI slices independently, overlooking the three-dimensional context of the brain. To address this, we propose a novel pipeline using YOLOv11 with a 2D+ approach, incorporating data from two orthogonal planes, comparing different combinations for axial, sagittal, and coronal. Our hypothesis is that this method will improve tumor localization by combining 2D anatomical views while maintaining computational efficiency. Our goal is to enhance clinical workflows, reduce manual annotation, and enable faster, more personalized treatment for glioma patients.

The workflow includes data processing, label generation, training, evaluation, and combining 2D labels for 3D masking. We used the BraTS2021 dataset [4], selecting the FLAIR modality for its sensitivity to edema around tumor cores. The 3D MRI volumes were converted into 2D JPEG slices, each paired with a segmentation mask for tumor identification. Bounding boxes were automatically computed from the ground truth segmentation masks. Maximum intensity projections in axial, sagittal, and coronal planes were used to derive 2D bounding boxes, and the final 2D+ bounding box was created by combining them. Labels in YOLO format were generated, and YOLOv11 was trained with an image resolution of 640×640 pixels. The model was tested on 2D slices, and predictions were mapped onto 2D+ space. Combined binary masks were used to crop the MRI volume, resulting in a tumor-focused sub-volume saved in NIfTI format. Model performance was evaluated using standard metrics such as precision, recall, and mean average precision (mAP) [5]. The overlap between predicted 2D+ masks and ground truth was quantified using the Dice Coefficient, demonstrating the effectiveness of the 2D+ tumor localization approach.

The YOLOv11 model trained on the 2D FLAIR MRI slices from the BraTS2021 dataset achieved effective tumor localization across the three anatomical planes. Quantitatively, the model obtained a mean precision and recall of 0.926 and 0.790 on the test set, reflecting the model’s ability to correctly identify tumor regions with a few false positives. After focusing on the 2D space detections, the combined 2D+ segmentation masks were compared against the ground truth. The resulting 2D+ masks enabled improved performance compared to individual 2D models, with the best results being obtained by combining Coronal and Axial views with a Precision of 100% and a Recall of 97%.

This study demonstrates that the YOLOv11 2D+ pipeline is a highly effective approach for brain tumor detection in MRI scans. The model achieved high mean precision, recall, and mAP values across the test set, indicating strong detection accuracy and robustness. Importantly, it maintained consistent performance across various tumor sizes and across the three MRI planes: axial, sagittal, and coronal. A major advantage of the YOLOv11 framework is its real-time inference capability, which, combined with its scalability to large datasets, makes it suitable for clinical environments where speed and efficiency are critical. The low rate of false negatives further enhances its potential utility in reducing diagnostic errors and radiologist workload. In addition to detection, the fusion of multi-view 2D predictions enables the generation of precise tumor-centered 2D+ sub-volumes. These focused regions improve the efficiency of downstream tasks, such as radiomics analysis, segmentation refinement, and treatment planning, by eliminating the need to process the entire MRI images [6]. The integration of this pipeline into clinical workflows could provide valuable support in neuro-oncology, particularly for early tumor diagnosis and intervention planning.

The YOLOv11 2D+ pipeline delivers accurate and efficient brain tumor detection across multiple MRI planes. Achieving a Dice Coefficient of 0.892 with minimal false positives, the model enables the generation of focused tumor sub-volumes that streamline downstream analysis. These results highlight the potential of YOLOv11 as a practical tool to support radiologists and enhance early diagnosis in clinical neuro-oncology workflows.

Alzheimer's disease (AD) is a multifactorial neurodegenerative disorder lacking effective curative therapies. Increasing availability of high-dimensional biomedical data has facilitated integrative strategies for early diagnosis and risk prediction. This study aims to develop and validate an integrative predictive modeling framework for AD by combining neuroimaging, genomic, and proteomic data using advanced machine learning and deep learning methodologies.

We analyzed data from the Alzheimer’s Disease Neuroimaging Initiative (ADNI), involving 469 participants classified into AD (n=153), mild cognitive impairment (MCI; n=157), and normal controls (NC; n=159). The dataset included structural magnetic resonance imaging (3D-T1 MRI), genome-wide single nucleotide polymorphism (SNP) genotypes, and plasma proteomic profiles. Radiomic features were derived through contourlet-wavelet transformations combined with gray-level co-occurrence matrix metrics. Genomic and proteomic features underwent dimension reduction via adaptive elastic net and group least absolute shrinkage and selection operator (LASSO) methodologies. Multiple classifiers—support vector machines (SVM), random forest (RF), gradient boosting machines (GBM), and extreme gradient boosting (XGBoost)—were compared. Additionally, a three-dimensional convolutional neural network (3D-CNN) was trained on voxel-level imaging data, with a hybrid 3D-CNN-SVM architecture subsequently constructed for enhanced diagnostic accuracy. Model performance was rigorously evaluated using 10-fold cross-validation and confirmed on an independent validation set, measuring accuracy, sensitivity, specificity, and area under the receiver operating characteristic curve (AUC). Model interpretability was assessed using SHapley Additive exPlanations (SHAP) and root mean square error (RMSE) analyses.

The hybrid 3D-CNN-SVM model exhibited superior performance in three-class diagnostic classification (NC vs. MCI vs. AD), achieving an accuracy of 92.11% ± 2.31%, significantly surpassing 2D-CNN and standalone machine learning models (P < 0.05). In binary classification tasks, the model achieved near-perfect accuracy (99.1%) and an AUC of 1.000 for discriminating AD from NC subjects. XGBoost classifiers built separately on genomic, proteomic, and integrated multi-omics data yielded accuracies of 92.9%, 98.8%, and 92.1%, respectively. Notably, the multi-omics XGBoost model augmented with demographic covariates (age, sex, education) provided optimal predictive performance for MCI-to-AD conversion (AUC: 0.945 [95% CI: 0.917–0.982], sensitivity: 0.976, specificity: 0.930). SHAP analyses identified critical protective biomarkers (e.g., rs7988039, apolipoprotein A-II) and risk-associated features (e.g., rs7191801, thrombospondin levels), correlating significantly with disease progression.

Our findings highlight the improved diagnostic accuracy afforded by integrative multi-omics modeling over single-modality approaches. The hybrid deep learning framework (3D-CNN-SVM) effectively captures neuroanatomical spatial patterns, while feature-informed XGBoost models capitalize on complementary information inherent to different omics data. Importantly, SHAP-driven interpretations provided biologically meaningful insights into AD pathogenesis, underscoring the clinical translational potential for personalized risk stratification.

This study establishes a robust, reproducible, and interpretable computational pipeline leveraging multi-omics integration for precise AD risk prediction. The synergistic combination of deep learning and interpretable machine learning with omics-derived biomarkers presents substantial promise for precision diagnosis and targeted early interventions in high-risk populations.

Ovarian cancer is a biologically heterogeneous disease, frequently diagnosed at an advanced stage with peritoneal dissemination. Identifying tumor compartments prior to histological analysis remains a major challenge, particularly for guiding biopsy and treatment planning [1]. Voxel-wise radiomics applied to high-field ex vivo MRI offers an opportunity to map tumor heterogeneity at high resolution [2]. We present preliminary results of habitat-based unsupervised clustering applied to texture-based radiomic maps derived from 9.4 T preclinical MRI of fixed ovarian tumor and peritoneal carcinomatosis samples.

Five ex vivo tumor samples were collected from two patients, including both primary ovarian tumors and peritoneal metastases. Samples were fixed in formalin for 48 hours and imaged in Fluorinert on a 9.4 T preclinical MRI scanner (fat-saturated gradient echo sequence, TR = 2000 ms, TE = 9.14 ms, flip angle = 60°, 4 averages). MRI preprocessing included N4 bias field correction, Gaussian filtering, and gray-level discretization (bin width = 25). Voxel-wise GLCM features were extracted using a 3×3×3 sliding window centered on each voxel within a segmentation mask [3]. To reduce feature redundancy and retain only informative descriptors of tumor heterogeneity, empirical entropy was computed within the segmentation mask for each GLCM feature. The most spatially diverse and non-redundant features were then selected for clustering. A KMeans clustering was then performed on the voxel-wise parametric maps to delineate distinct imaging habitats. Habitat maps were visually assessed and compared to digitized histological sections, which were aligned to the MRI acquisition plane and segmented using QuPath based on expert annotations.

Based on the entropy analysis, four GLCM features were retained for clustering: Difference Entropy, Correlation, Inverse Difference (Id), and Inverse Variance (Figure1) Unsupervised clustering was initially performed with KMeans (K=2), but this configuration consistently yielded a marginal habitat confined to the edges of the segmentation mask, likely reflecting an edge effect introduced by the radiomic sliding window. To overcome this, clustering was repeated with K=3, which revealed two spatially coherent internal habitats and a third peripheral one. This marginal habitat was then merged with its adjacent internal zone based on spatial continuity, resulting in a final binary habitat segmentation. The two resulting radiomic habitats showed visual strong correspondence with histological compartments across all five samples, with one region aligning with densely cellular tumor areas and the other with stromal or low-cellularity regions (Figure 2).

This work demonstrates that unsupervised clustering of voxel-wise GLCM features from high-field (9.4 T preclinical) ex vivo MRI can delineate distinct imaging habitats within ovarian cancer tissue. Entropy-based feature selection enabled the identification of spatially diverse and complementary texture parameters. These features captured complementary textural information, reflecting both heterogeneity (Difference Entropy), structural organization (Correlation), and local uniformity (Id and Inverse Variance). After correcting for edge-related artifacts, the resulting radiomic habitats exhibited reproducible spatial organization and strong concordance with histological compartments across all samples. The next step is to spatially register and overlay these MRI-derived imaging habitats with QuPath-based histological segmentations. This alignment will enable voxel-level tissue classification and support the development of supervised models for automated tissue-type prediction.

These preliminary results suggest that radiomic clustering based on GLCM features from high-field ex vivo MRI can reveal biologically meaningful imaging habitats in ovarian cancer tissue. Future work will extend this approach to quantitative MRI sequences (such as T1 and T2 mapping), integrate spatial transcriptomics, and develop predictive models toward robust radiogenomic signatures for comprehensive tissue characterization.

The UK Biobank (UKB) is collecting multimodal brain MRI scans of 100,000 participants using the same acquisition protocol on identical scanners and extracting imaging derived phenotypes (IDPs) with a standardised pipeline [1-2]. However, confound effects are still present and need to be carefully addressed. In previous work [3], it has been shown that when the T2-FLAIR image is absent, FreeSurfer [4-5] IDPs are affected. While only around 2% of the participants in UKB are missing T2-FLAIR images, this accounts for up to ~14% variance explained and ~7% of unique variance explained in some IDPs [3], which is substantial enough to cause significant confounding when investigating associations with non-imaging variables (nIDPs). In this work we sought to investigate the prevalence and size of this T2-FLAIR “batch effect” confound across structural IDPs. We then assessed strategies for mitigating this effect, comparing established harmonization approaches. Finally, we investigate the impact of not correctly addressing this bias in downstream analysis on the associations between affected IDPs and non-imaging variables (nIDPs).

We first looked at the magnitude of the T2-FLAIR batch effect across 1432 structural IDPs, using Cohen’s d and examining the ratio of variance between the two groups (T2-FLAIR available vs no T2-FLAIR). We also created age nomograms for IDPs that are known to change with age to show potential downstream effects not addressing this bias. We then compared several methods addressing this bias while also handling four other standard confounds (age, sex, head position and head size scaling). We first used multiple linear regression of the four confounds and a binary indicator for the missing T2-FLAIR. We then used a modified form of ComBat, which instead of preserving the other confound variables, uses the predicted coefficients to remove their effects from the data at the point at which we apply the batch correction [6-7]. Additionally, to test Combat's advantages over regression, we tested our modified method in two ways: (A) turning off ComBat’s batch rescaling of the IDPs (leaving just the mean shift or bias correction); (B) modifying ComBat to only pool prior distributions for variance scaling and mean shift within classes of IDPs, and not across all different classes (e.g., cortical thickness vs surface area). We validated the effectiveness of these approaches against each other and for when no correction had been applied by comparing Cohen’s d and within-group variances. Finally, we assessed some potential downstream analysis implications of not addressing this effect by computing the Fisher-transformed correlation coefficients with a set of over 18,000 nIDPs, comparing our different test cases against each other and against the case where no correction for missing T2-FLAIR had been applied, using Bland-Altman plots.

We found a strong negative effect size between the two groups in over 250 cortical volume and thickness IDPs. In the most affected IDPs, the ratio of variance between the groups with and without T2-FLAIR was as high as 2.8, with the difference being most prominent in FreeSurfer cortical thickness IDPs (Figure 1). When using ComBat, the effect size (mean shift bias) is reduced to below 0.1 for all IDPs, the variance ratio is closer to 1 and the mean shift across ages is corrected (Figure 2 and 3). When using simple regression, the effect size is approximately zero for all IDPs, but the ratio of variance is uncorrected. When looking at associations with nIDPs, ComBat gives stronger correlations than no handling of missing T2-FLAIR or handling via simple regression (Figure 4). When comparing ComBat pooling its priors only within IDP groups, and without the scaling term, we found that ComBat run on all IDPs at once and with the variance scaling term included showed the best reduction of bias shift, a variance ratio closest to 1, and marginally stronger associations with nIDPs.

The omission of T2-FLAIR causes a strong bias in the mean and variances of over 250 IDPs. We have shown ComBats advantage over simple regression in its ability to adjust the variance of the data. Through comparisons of the correlations with nIDPs, we have shown that using ComBat to handle both confound and batch correction is beneficial when compared to simple regression. Despite different groups of FreeSurfer IDPs not being affected in the same way by the omission of T2-FLAIR, pooling ComBat priors from all IDPs seems to be beneficial. Additionally, when the bias affects more than just the means, the scaling correction is a beneficial addition over simple regression.

To conclude, treating the missing T2-FLAIR effect as a confound and addressing it using regression doesn’t sufficiently correct for its effects. Researchers using FreeSurfer derived IDPs for their own data, or using UK Biobank IDPs should be aware of this and consider using ComBat to handle both the T2-FLAIR effect and other confound corrections.

The consistency of fMRI measurements upon repetition, known as test-retest reliability, is essential for interpreting the validity of neuroimaging results, as measures must be reliable to have statistical power. However, quantifying the reliability of resting-state fMRI is complex. Previous research has indicated poor reliability for univariate functional connectivity measures at the edge level [1]. In contrast, multivariate metrics reflecting whole-brain patterns tend to demonstrate greater reliability [2]. Additionally, the duration of fMRI scans has been found to influence reliability estimates [3]. Despite significant efforts to evaluate test-retest reliability in resting-state fMRI, a comprehensive comparison of various reliability metrics across intra- and inter-session data, while systematically varying scan lengths and sample sizes, is still required. We aim to investigate the intra- and inter-session test-retest reliability of both univariate and multivariate resting-state fMRI functional connectivity metrics under different experimental conditions.

We analysed minimally pre-processed resting-state fMRI data from 100 unrelated healthy adults (54 females, 46 males, ages 22–35) provided by the Human Connectome Project. Data included two sessions (REST1 and REST2) and two runs within the first session (REST1_LR and REST1_RL). Following high-pass filtering and spatial smoothing using FSL, we assessed test-retest reliability both within and between sessions. We systematically varied the number of volumes (300, 600, 900, 1200) and sample sizes (10, 20, 50, 100). Both a voxel-based and a node-based analysis were conducted: 1) Group Independent Component Analysis (ICA) with 50 components was conducted to identify canonical resting-state networks (RSNs), and the spatial similarity between these ICA-derived networks and template RSNs was quantified using the Dice coefficient [4]; and 2) Functional connectivity matrices were computed using Pearson correlation on data that was parcellated into 400 regions based on the Schaefer atlas. We evaluated one univariate metric, the average intraclass correlation coefficient (ICC) across all edges within each RSN [5], as well as two multivariate metrics: fingerprinting, which compares the correlation between connectomes [6], and discriminability, which evaluates the distance between connectomes [7].

At the group level, the spatial similarity between ICA-derived networks and canonical RSN templates remained consistent across all seven RSNs, regardless of the number of volumes analysed (Figure 1A). Spatial reproducibility of ICA-derived networks, when comparing runs within and between sessions, varied by RSN (Figure 2A). The Visual Network (VN), Sensorimotor Network (SMN), and Frontoparietal Network (FPN) demonstrated the most consistent Dice coefficients across different scanning lengths, particularly when considering 50 and 100 subjects. At the individual subject level, mean spatial similarity increased with longer scan durations and remained relatively stable across varying numbers of subjects (Figure 1B). Similarly, spatial reproducibility within and between sessions improved with increasing scan length, while the number of subjects had a less significant impact (Figure 2B). For connectivity matrix-based metrics, edge-level ICC reliability improved with longer scanning durations, transitioning from poor to fair reliability (ICC > 0.4) (Figure 3A). The FPN, SMN, Dorsal Attention Network (DAN), and VN generally showed the highest edge-level ICC values. Additionally, the multivariate metrics, differential identifiability and discriminability, both increased with longer scanning lengths (Figure 3B).

This study demonstrates the impact of scan length and sample size on the test-retest reliability of various fMRI metrics. While group-level ICA for network identification remains robust across scan lengths, the reliability at the individual level improves with longer scan durations. Edge-level functional connectivity shows modest improvements with extending scanning times. In contrast, multivariate whole-brain metrics demonstrate greater reliability with increased scan length, suggesting that whole-brain patterns tend to be more stable than individual connections.

We show strong variations of the test-retest reliability across various resting-state fMRI functional connectivity metrics, across scanning durations and sample sizes.

Aceruloplasminemia (ACP) is a rare neurodegenerative disorder characterized by a complete loss of ceruloplasmin ferroxidase activity, leading to systemic iron overload in the brain, liver, and pancreas [1]. While MRI-based iron mapping, using R2* or Quantitative Susceptibility Mapping (QSM) is widely applied, microstructural iron quantification faces unresolved challenges due to the multiple susceptibility sources which contribute to the measured MR signal [2]. Microstructural iron quantification using MRI microscopy images with a resolution in the µm scale has been extensively studied, however reconstructing magnetic contributions from images in the mm scale, as acquired in a clinical setting, remains complex with inherent mathematical limitations in certain regimes [3]. This study leverages the Static Dephasing Theory (SDT) [4], a framework modeling transverse relaxation, to quantify iron accumulation in ACP brain tissue. SDT describes relaxation in the presence of magnetized objects (e.g. iron-rich cells, blood vessels) that induce local magnetic field inhomogeneities. The theory applies in the regime where these magnetic perturbers are sufficiently large, such that diffusion has negligible effect on the relaxation process. We demonstrate SDT’s capacity to extract microstructural biomarkers (e.g. inclusion volume fractions, absolute magnetic susceptibility) in ACP, advancing beyond conventional R2* or QSM.

One patient with ACP and a matched healthy control were imaged on a 3T MRI system (MAGNETOM Skyra, Siemens Healthineers, Erlangen, Germany). The following sequences were acquired: T1-weighted MPRAGE for anatomical segmentation, multi-echo GRE for QSM and R2* mapping, and dual-echo SPACE for R2 mapping. T1-weighted images were registered to GRE space using FLIRT, and segmentation was performed with FSL FIRST. R2* maps were computed using ARLO, and QSM reconstruction included ROMEO phase unwrapping [5], V-SHARP background field removal [6], and iLSQR dipole inversion [6], with susceptibility values referenced to the CSF. Using SDT, we modeled the signal decay to extract additional microstructural parameters, such as volume fraction and absolute susceptibility of magnetic inclusions, which are not accessible with standard approaches. The static dephasing signal was numerically evaluated and fitted to a voxel subset in the basal ganglia using the L-BFGS-B algorithm with biophysical motivated constraints on the microstructural parameters.

Figure 1 provides an overview of the R2* and R2 maps in the patient and healthy control. As shown in Figure 2, R2* values are markedly elevated in all basal ganglia regions of the patient, with the highest in the putamen (R2*= 125.3 ± 42.1 s⁻¹), corresponding to an average 3.14-fold increase over the control. In contrast, R2 values increase only moderately, with a mean factor of 1.57. Microstructural parameters from SDT are summarized in Table 1 and compared with R2 values from the SPACE sequence (R2SE) and QSM-based susceptibilities (X_QSM). Figure 3 shows the fitting curves for the putamen. In the patient, SDT-derived absolute susceptibilities (|X_SDT|) range from 0.176 ± 0.096 ppm in the pallidum to 0.226 ± 0.094 ppm in the putamen, consistently exceeding QSM values, which range from 0.01 ± 0.16 to 0.24 ± 0.33 ppm. In the control, this overestimation is even more pronounced. Notably, |X_SDT| values do not significantly differ between patient and control. The SDT model also revealed elevated volume fractions (η) of magnetic inclusions in the patient, exceeding 10% in all regions and peaking at 13.5 ± 8.0% in the putamen, compared to an average of 2.95% in the control. In nearly all regions, SDT yielded a lower mean squared error (MSE) than the mono-exponential model, except for the control’s caudate.

Our findings suggest that the altered transverse relaxation observed in the basal ganglia of ACP patients arises primarily from the inhomogeneous distribution of magnetic susceptibility sources, rather than solely from elevated iron concentrations. The notable difference in R2*/R2 ratios between groups indicate that static dephasing is the dominant signal decay mechanism in ACP. This is supported by the post mortem observation of clustered ferritin-bound iron aggregates, generating localized susceptibility gradients [1, 2]. Volume fractions over 10% in regions like the putamen highlight clustered iron buildup in ACP. Similar vascular contributions in patients and controls make vascular changes unlikely to explain the pronounced contrast. These findings support ferritin aggregation as the main source of susceptibility alterations in ACP.

This study demonstrates the utility of SDT-based modeling of iron related microstructural alterations in ACP. Beyond mapping the spatial distribution of ferritin-bound iron, this method allows quantification of microstructural biomarkers directly from the MR signal decay. Such advancements will allow new mechanistic insights into ACP pathogenesis.

Alzheimer's disease (AD), an irreversible neurodegenerative disorder affecting over 55 million people worldwide[1], presents a critical public health challenge intensified by aging populations. Advances in neuroscience and radiomics now enable the systematic tracking of AD progression through conventional MR images[2]. Recent studies have increasingly focused on evaluating the impact of MRI field strengths on diagnostic performance, particularly modeled as classification tasks[2-4]. The radiomics features extracted from MR images, correlating with hippocampal atrophy and cortical thinning, may be considered valuable biomarkers reflecting pathological and functional changes. However, a critical gap persists in understanding how individual MRI-derived features behave across field strengths. In brain imaging, MR images are acquired using different manufacturers, magnetic field strengths, and acquisition parameters. These variations can significantly affect further radiomics studies, especially different magnetic field strength. In this study, our objective was to assess the robustness of specific radiomic features across different MRI field strengths and evaluate their diagnostic value for AD detection.

This retrospective study analyzed 253 subjects from the ADNI database (https://adni.loni.usc.edu) with paired 1.5T and 3T MRI scans . Of these, 191 subjects were included because their scans were conducted within a 3 month period: 41 Alzheimer’s disease (AD) patients, 88 mild cognitive impairment (MCI) patients, and 62 cognitively normal controls (NC) (age 74.67 ± 7.21 years; 95 males/96 females). All scans utilized standardized T1-weighted sequences with harmonized parameters across scanners/sites under quality control at the Mayo ADIR lab. Bilateral hippocampi were segmented based on HippoDeep method[5] for 1.5T and 3T scan respectively. Subsequently, radiomics features were extracted from the segmented hippocampi using Pyradiomics[6]. To evaluate the reproducibility and robustness of the extracted features across 1.5T and 3T scans, the nonlinear concordance correlation coefficient (NCCC) and the Wilcoxon signed-rank test (WSR) were used. The acceptable stable features (NCCC > 0.6) were used for classification tasks to distinguish between NC and MCI, as well as between MCI and AD. Rather than using the features that passed the Wilcoxon signed-rank test, the ten most informative features were selected based on mutual information scores. A Random Forest classifier was trained on the training set (80%), and its performance was evaluated on the testing set (20%). To enhance robustness and reduce the risk of overfitting, five-fold cross-validation was performed within the training set during model development.

Figure 1 shows representative hippocampal segmentation from NC, MCI, and AD subjects' MRI at both 1.5T and 3T field strengths. Among the 120 radiomic features, 27 features displayed acceptable concordance between 1.5T and 3T (NCCC > 0.6). Most of the features were shape-related, as shown in Figure 2. Figure 3 shows the classification performance based on area under curve (AUC) for distinguishing NC-vs-MCI and MCI-vs-AD groups at both 1.5T and 3T field strengths.

In this work, several stable features were identified that were not significantly affected by field strength and were able to preserve hippocampal characteristics across different field strengths. Using stable features, four Random Forest classifiers were constructed to successfully classify subjects into NC, MCI, or AD. These models achieved above 80% area under curve (AUC) for MCI-vs-NC, i.e. strong discrimination. However, differentiating MCI from AD remains challenging, with 71% and 76% AUC, respectively for the 1.5T and 3T MRIs. Potential reasons for this are that i) MCI encompasses diverse range of microstructure changes, making it a complex condition to classify, and ii) subtle features captured by NCCC are insufficient to detect changes in MCI progression. Therefore, future research may need to incorporate more refined MCI stratification strategies and employ enhanced feature techniques for comparing feature robustness to improve the classification performance between MCI and AD.

The common radiomics features across 1.5T and 3T demonstrate potential for distinguishing NC, MCI and AD. Additionally, those stable features truly possess generalizable diagnostic value for early Alzheimer's disease.

Aging has been known to show changes in cerebral perfusion (Liu et al. 2012) that has led to many large consortia acquiring ASL data to non-invasively map perfusion. However, these studies have acquired data at very low-spatial resolutions (≥3.5mm in-plane and 5-7mm slice thickness). Whilst this may be sufficient to map cortical perfusion and age- and disease-related perfusion changes in at the scale of whole brain areas, it is insufficient to resolve (Kashyap et al. 2024) smaller, deeper brain areas such as hippocampus, thalamic nuclei that are known to be impacted in disorders such as Alzheimer’s (Rodriguez et al. 2000) or Parkinson’s (Kimura et al. 2011). For this ongoing cross-sectional study, we optimised a high-resolution ASL protocol (2.5 mm iso) to acquire CBF maps in clinically-feasible times (< 6 min) at 3 T, and demonstrate its utility to capture perfusion-changes in thalamic nuclei in a cohort of cognitively normal older adults.

2.5 mm isotropic 3D-GRASE pCASL (TR/TE=3.93s/20ms, GRAPPA 3, PF=6/8, Label=1.8s,PLD=1.8s,TF=12,Seg=4,EPI Factor=27,BW=2480Hz/Px, TA=5:42 min) was acquired in 35 healthy older adults (24F,11M) aged 56.0 to 84.5 years (mean ± SD = 71.9 ± 8.1 years) on a Siemens Prisma 3T (Siemens Healthineers, Erlangen). Data were processed using oxasl (Chappell et al. 2023) and THOMAS (Su et al. 2019) was used to obtain delineations of thalamus and its nuclei. ANTs (Tustison et al. 2021) was used for all registrations and cohort template building.

Group average THOMAS parcellation (Fig 1a) and corresponding volumes (Fig 1b) are consistent with Su et al. 2019. Perfusion-weighted image and CBF maps (Fig 1c) show observable differences in the thalamus (white outline) and the mean CBF (Fig 1d) differed markedly across thalamic subregions. The habenula exhibited the lowest perfusion (~35 ml/100 g/min), whereas the anteroventral and mediodorsal nuclei showed the highest (~73 ml/100 g/min). Most nuclei fell within one standard deviation of the global thalamic mean (~56 ± 23 ml/100 g/min) and show negligible lateralization. There was a non‐significant mild positive correlation between nuclei volume and mean CBF (r(10) = 0.50, p = 0.098, R² = 0.25), indicative of a trend toward higher perfusion in larger nuclei. A non-significant inverse relationship between age and perfusion (slope ≈ –0.15 ml/100 g/min per year) can be observed (Fig 1e) consistent with expectations in healthy aging. Region‐specific trends are similar between sexes, with females having consistently higher CBF by ~5–15 ml/100 g/min (Fig 1f).

We demonstrate the feasibility of a 2.5 mm isotropic 3D-GRASE pCASL protocol at 3 T to resolve perfusion differences across individual thalamic nuclei in a clinically acceptable scan time (< 6 min) in 35 cognitively normal older adults. We show an approximately two‐fold range in baseline CBF across nuclei (Hb ~ 35 ml/100 g/min to AV and MD ~ 73 ml/100 g/min). The observed perfusion differences indicate differing cytoarchitecture and metabolic demands of thalamic subregions. These results, part of an ongoing clinical study, establishes a normative baseline to compare subregional thalamic dysfunction in early diagnosis and monitoring of disease progression in Alzheimer’s and Parkinson’s disease.

We show high-resolution ASL at 3T overcomes the limitations of conventional ASL to non-invasively map perfusion and resolve CBF changes in small, deep grey-matter structures such as thalamic nuclei, and demonstrate the capability to acquire them in clinically-feasible times.

Ensuring the reliability of metrics derived from Magnetic Resonance Imaging (MRI) data is essential for both clinical and research applications. Scanner relocation to a new facility poses a potential challenge to the stability of MRI-derived data [1]. This study assessed the reproducibility of multispectral neuroimaging measures including diffusion, resting-state networks (rs-fMRI), task-based fMRI [2], and morphometry acquired before and after relocation of a 3T MRI scanner, in the absence of any software or hardware upgrades.

Six healthy volunteers completed 4 MRI sessions in a test-retest design before and also after scanner relocation (3T Siemens Prisma, 64 channel head-neck receive coil). Each session included: T1w Multi-Echo MPRAGE (meMPRAGE), Multishell DWI, UMN CMRR multi-echo EPI for rs-fMRI, single-echo EPI for task-based fMRI. Assessed metrics were: segmentation-based cortical and subcortical gray matter (GM) volume [3]; Mean-Signal Kurtosis (MSK; MS Diffusion Kurtosis Imaging, MS-DKI [4]) in GM and white matter [5] regions of interest (ROIs); strength and topology of the rs-fMRI Default Mode Network (DMN); Euclidean distance between functional peaks in ROIs based on contrasts mapping auditory areas and ventral/dorsal visual areas. Data were compared across the four sessions in terms of: absolute values, relative change, and reliability in all the considered metrics.

Brain morphometry reproducibility (Figure 1): Two-way ANOVA on T1w structural data showed no significant results both for main effects (Session, Site) and their interaction, in any of the considered ROI (pFDR=NS). Relative changes remained well below the 5% threshold for within- and between-sites test-retest. One-way ANOVA on relative variations was non-significant. ICC(3,1) revealed excellent reliability (average ICC=0.98; CI(95%)=0.94-0.99). Brain microstructure reproducibility (Figure 2). Relative MSK changes were below 3% across Sites and Sessions, with excellent reliability (ICC(3,1)=0.99; CI(95%)= 0.98-0.99). No site effect was observed in any ROI (all padj>0.3). Intrinsic functional connectivity reproducibility (Figure 3). Resting-state connectivity analysis on average Z-score extracted from DMN revealed no significant differences for both main effects and their interaction in a two-way ANOVA. A further one-way ANOVA on relative changes was non-significant. Test-Retest measurements assessed with ICC(3,1) showed limited variability with an estimate of 0.99 (CI(95%)= 0.9-1), in good agreement with literature [6]. Spatial distribution of DMN was also tested across Sites and Sessions: non-significant differences were reported with TFCE-corrected non-parametric permutation testing. Task-based fMRI reproducibility (Figure 4). The Euclidean distance between two selected peaks in the ROI analysis—performed separately for the auditory–visual and visual–auditory stimulus contrasts—revealed no significant main effects (Site, Sessions) or interaction (two-way ANOVA, p=NS). Moreover, the two one-way ANOVAs (one for each contrast) on the relative change of the Euclidean distance were not significant. On average, the relative variation of the Euclidean distance between peaks remained below the 5% threshold for both within- and between-sites comparisons. Finally, the two ICC(3,1) performed on the Euclidean distance in both contrasts indicated a good reliability, with estimates of 0.60, (CI(95%)=0.3-0.8) and 0.70 (CI(95%)=0.5-0.9), respectively.

By systematically comparing pre- and post-relocation data within the same healthy volunteers, we aimed to determine if scanner relocation introduces biases in the considered metrics. Our data show high consistency across all examined measurements. Despite technical changes, such as hardware calibration, and differences in environmental factors like room characteristics and ambient conditions, we did not observe any meaningful variation in the results. Test-retest reproducibility within- and across-site locations were not significantly different.

Scanner relocation had minimal impact on our neuroimaging measurements: structural measures varied <5% (ICC = 0.98), diffusion MSK <3% (ICC = 0.99), task-based fMRI activation peak distances <5% (ICC = 0.60-0.70). A high overall reproducibility (ICC = 0.99) was found even for resting-state DMN topology. These results confirm the robustness of MRI neuroimaging measures post-relocation.

Simultaneous electroencephalography (EEG) - fMRI acquisitions are used to investigate brain activity due to the highly complementary characteristics of the two neuroimaging modalities [1]. EEG-fMRI studies of the sensorimotor network (SMN) are of particular interest for applications in the neurorehabilitation of motor function, such as neurofeedback (NF) during motor imagery (MI) [2]. However, studies comparing EEG fluctuations with concurrent SMN fMRI activity have reported inconsistent findings, typically showing only very low-amplitude temporal correlations [3]. We hypothesize that this may partly stem from fluctuations in these correlations over seconds to minutes and also across different conditions. Here, we investigate the within-subject dynamics of temporal correlations between SMN fMRI and concurrent EEG spectral power across both task and rest conditions.

Simultaneous EEG-fMRI data were recorded from 15 healthy subjects in 2 sessions ~2 weeks apart. In each session, subjects participated in two ~17min MI tasks (Graz and NeuRow) and a 10min rest scan with eyes open (Rest). fMRI data were acquired on a 3T Siemens system using 2D EPI (TR/TE=1260/30ms, 60 axial slices, 2.2mm iso). EEG data were collected using a 32-channel MR-compatible BrainAmp system, with 31 EEG channels and 1 ECG channel, synchronized with the fMRI scanner. Regarding MI tasks, after standard preprocessing, a GLM analysis was conducted to identify the SMN and extract its average BOLD signal. A BOLD time series representative of the spontaneous trial-by-trial fluctuations (TBT) was also obtained by regressing out the task-related effects. For Rest, after standard preprocessing and denoising, a group-level ICA was applied to identify the SMN, which was back-projected to individual subjects to obtain subject-specific BOLD time series. After artifact removal and standard preprocessing, EEG data underwent time-frequency decomposition using Morlet wavelets (3s temporal, 1Hz frequency resolution) to obtain the relative power across frequencies and time points for each channel. These power time series were downsampled to match the temporal resolution of the BOLD signal using an anti-aliasing low-pass filter. For each condition (Graz-TE, NeuRow-TE, Graz-TBT, NeuRow-TBT, Rest), subject, and session, sliding windows of 36 TRs with ~90% overlap were applied to EEG and fMRI data. For each sliding window, Pearson correlation was computed between the SMN fMRI time series and EEG power across 40 frequency bins (1-40Hz) and 31 channels, considering 9 time lags (0-8 TRs~10s). The time lag yielding the strongest correlation was selected for each condition, subject, session, channel, frequency bin, and sliding window. K-means clustering was used to identify recurrent correlation patterns, exploring cluster numbers from k=2 to k=8, with further analysis revealing 4 distinct states.

Fig.1 displays the centroids of the 4 states (A, B, C, and D) found in the EEG-fMRI correlation dynamics for each condition. Fig.2 shows time-dependent EEG-fMRI correlations for Graz-TBT, averaged across channels and concatenated across subjects and sessions for each connectivity state. Fig.3 presents the Euclidean distances between EEG-fMRI correlation states for different conditions. States A and B demonstrated consistent patterns across all conditions, as reflected by their lower Euclidean distances between conditions. In contrast, states C and D exhibited greater variability, with state C showing the largest distances in most cases. Fig.4 illustrates the temporal dynamics of the states’ parameters, including dwell time and fractional occupancy. A two-way repeated measures ANOVA revealed a significant main effect of state for both dwell time (F=2.97, p=0.04) and fractional occupancy (F=4.1, p=0.01). Additionally, a significant interaction between state and condition was observed for both measures (dwell time: F=7.3, p<0.001; fractional occupancy: F=3.4, p<0.001). Post-hoc analyses, adjusted with Bonferroni correction for a family of 40, were conducted to explore the significant (p<0.05) 2-way interaction identified between state and condition. Differences in dwell time between conditions were found for states A, B and C. No significant effects were found for fractional occupancy in the post-hoc comparisons.

While states A and B displayed symmetric correlation patterns across a broad frequency band above 10 Hz, states C and D showed symmetric correlation patterns centered around the alpha band near 10 Hz. The correlation patterns remained consistent across time windows for each connectivity state, suggesting stable EEG-fMRI coupling within each state. Dwell time results show that the stability of each state varies depending on the conditions.

Our results reveal high variability in EEG-fMRI correlations in the SMN over time and across different task and rest conditions, offering a potential explanation for the consistently low-amplitude correlations reported in the literature.

Correlation Tensor Imaging (CTI) is an advanced diffusion MRI method based on Double-Diffusion-Encoding (DDE) that separates diffusional kurtosis into isotropic (Kiso), anisotropic (Kaniso), and microscopic (μK) components, offering greater specificity for brain microstructure characterization [1]. While CTI has shown strong potential in preclinical studies [2], and was recently demonstrated in humans [3], current protocols require long acquisitions (~50 min), limiting clinical feasibility. Accelerating CTI will drastically reduce acquisition time but also exacerbate signal-to-noise ratio (SNR) losses, making effective denoising crucial. PCA-based methods such as Marčenko-Pastur PCA (MP-PCA) [4] and Threshold PCA (TPCA) [5] have shown promise for denoising diffusion MRI data without compromising microstructural information. While MP-PCA denoising classifies principal components based on the Marčenko-Pastur distribution, assuming spatially uncorrelated noise, TPCA incorporates a prior estimate of the noise variance to establish a threshold for noise component removal, enhancing robustness in scenarios with spatially correlated noise. This study evaluates, for the first time, the effects of MP-PCA and TPCA denoising on CTI-derived metrics in healthy human brains.

Population & MRI Acquisition: We used CTI data acquired from 8 healthy young volunteers, as described in a previous 3T study [3]. Preprocessing and CTI: Three CTI preprocessing pipelines (P1–P3) were implemented, differing only in their PCA denoising approach: P1 – no denoising; P2 – MP-PCA (MRtrix3) [4]; P3 – TPCA (Dipy) [5]. After denoising, all pipelines included the same subsequent corrections in the following order: Gibbs ringing (MRtrix3), geometric distortions and eddy currents (FSL), signal drift [6], and bias field (MRtrix3). CTI metrics were fitted and estimated within different brain regions of interest (ROIs). Statistical Analysis: To assess the impact of denoising across different processing pipelines, we compared three key metrics within ROIs: 1) the mean CTI estimates, 2) the within-ROI variability of CTI values, and 3) the percentage of voxels with biologically implausible (i.e., negative) CTI fits. Statistical analysis was conducted using within-subject repeated measures. A Friedman test was first applied to identify overall effects of the pipeline. When significant, this was followed by right-tailed Wilcoxon signed-rank tests for pairwise pipeline comparisons. P-values were adjusted for multiple comparisons using the False Discovery Rate (FDR) method.

Figure 1 shows sample subject results from the three pipelines on CTI metrics. Denoising improvements are most visible on μK. At the group level, there were no significant differences across pipelines on CTI metrics averaged in the ROIs. When looking at within-ROI variability and percentage of negative CTI fits, we observed significant improvements with denoising, in particular with TPCA (Table 1). Figure 2 shows how voxel-wise variability of CTI metrics is systematically and significantly reduced within the various ROIs considered when moving from P1 to P2 to P3, especially for μK (p < 0.01). In addition, we found similar effects of denoising improvements when looking at the percent of negative CTI model fits (biologically impossible values) within the various ROIs, especially for μK (Figure 3).

Our main findings are twofold. First, PCA-based denoising significantly reduces within-ROI variability of CTI metrics and decreases the number of voxels yielding biologically implausible model fits. Second, these denoising effects are markedly stronger when using TPCA compared to MPPCA. Both effects were consistently observed across gray and white matter regions in healthy young brains. CSF showed the strongest denoising improvements in terms of reduction of negative fits, maybe related to the fact that Rician denoising was not included. These preliminary results highlight the value of incorporating PCA denoising strategies into the human CTI protocol at 3T to enhance data quality and reliability, in particular TPCA.

While these findings are promising, further work is needed to further characterize CTI denoising optimization for clinical and research applications. Future studies will investigate the additional impact of Rician bias correction within the CTI framework on a larger sample, as well as assess how these denoising approaches influence test-retest variability, an important marker of reproducibility.

A fundamental and hitherto unsolved challenge in diffusion MRI tractography is to uncover the spatial distribution of orientations within an imaging voxel from the voxelwise fibre orientation distributions (FODs). Streamlining in tractography attempts to solve this problem through a combination of interpolation and heuristics [1], such as penalising sharp turns in the streamlining. Although this approach has been successful over the years in mapping large coherent bundles, tractography notoriously still suffers from "bottleneck" issues where the FOD can ambiguously represent different underlying orientation configurations [2]. The problem is that once we've built an FOD, it is too late to recover the underlying pointwise orientations. Here we propose a potential paradigm shift: rather than estimating pointwise orientations through streamlining after fitting FODs, we propose to model the underlying pointwise orientations directly in a forward model of the data that bypasses the need for fitting FODs (Fig 1A). Pooling the data across neighbourhoods of voxels makes this model tractable (invertible). We present a preliminary version of the model in simulated and real data, and suggest how this type of approach can be used for joint modelling of dMRI and polarised light imaging (PLI) data from the same tissue, where the PLI further constrains the model.

At the core of the forward model are simple multi-layer perceptrons (MLPs, which in these experiments have 4 layers, 256 neurons per layer, tanh activations, linear output). The MLPs take as input the (x,y,z) coordinates of any point in space (not necessarily at the centre of a voxel), and output either a 3D vector (for modelling the orientation vector field) or a scalar (for modelling e.g. the diffusion coefficient). To enable modelling of both high and low frequencies in the spatial domain, we insert a Fourier Feature embedding layer between the input and the MLPs [3]. The outputs of the MLPs (for the vector and scalar fields) are then used within forward models of the data (Fig 1B). For example, to predict diffusion data, sample positions from within a voxel are fed through the MLPs to generate orientations and scalars, then fed through the diffusion model equation for a stick given a set of bvals/bvecs, and then averaged over the voxel. The predicted signal is then compared to the measured diffusion data to calculate a mean squared error loss. This is summed over a neighbourhood of voxels to calculate the overall loss used for training the MLPs. Other data, such as polarised light imaging (PLI), can be similarly predicted from the orientation vector field, enabling the joint modelling of multiple modalities (optional extra). We call this method SPOT (Spatial Pointwise Orientation Tracking). SPOT was evaluated using simulated diffusion data from an orientation vector field (here 2D) which we aimed to recover. For comparison with interpolation, the data was resampled onto a 100x100 grid using splines of order 3 (with Scipy's map_coordinates), and a tensor was fitted in each pixel to obtain a principal orientation per pixel. Next, we assessed the impact of jointly fitting a single dMRI voxel alongside one or two slides of high-resolution PLI intersecting the voxel. Finally, we show some early results fitting SPOT to in vivo dMRI data (60 directions, 2mm isotropic, b=1k).

Preliminary results are shown in figures 2-4. In 2D simulations, we compare SPOT to spline interpolation (Fig 2) for different grid sizes (1x1, 2x2 and 4x4) to demonstrate the effect of including neighbourhood information. SPOT outperforms interpolation in terms of the resolved pointwise orientations and the FODs. When jointly fitting to MRI and PLI (Fig 3), SPOT manages to recover the underlying pointwise orientations from a single voxel (no neighbourhood information) even with a small amount of PLI as additional constraint . Results comparing SPOT to constrained spherical deconvolution (CSD) in in vivo data (Fig 4) show how FODs with crossing fibres resolve into streamlines bending onto the cortex in SPOT.

SPOT is still very much under development. Future work includes: investigation of the effect of spatial resolution, b-values, inclusion of multiple modalities, and size of the spatial neighbourhood. Current implementation is limited to a small neighbourhood, and further work is needed to exploit the pointwise vector field for mapping long-range connectivity.

SPOT is a super-resolution, multi-modal technique that attempts to recover the pointwise orientation (and other scalar) field(s) directly from voxelwise dMRI data.

Graph theoretical analysis of brain networks is a powerful method for quantifying topological features that may help differentiate between healthy and disease-affected brains. During analysis of networks based on resting-state functional MRI (rs-fMRI), nodes are defined as parcellation-derived regions of interest (ROIs) and edge weights as temporal correlation coefficients. [1] Analyzing neuroimaging-derived networks is hindered by the multiplicity problem, the unclear nature of data distribution and having insufficient prior knowledge on the location of effects of interest or the core connectome structure. [2,3] We investigated rs-fMRI derived functional brain networks of controls and patients diagnosed with epilepsy to better understand the core structure and topology of these networks and also the effects of thresholding on graph theory metrics. We aimed to identify regions that can be considered hubs in healthy controls and investigate how hub-identification depends on network thresholding and the fineness of the ROI-resolution.

Rs-fMRI of 28 healthy controls and 34 patients diagnosed with epilepsy were acquired on the 3T Philips Ingenia scanner of Semmelweis University Medical Imaging Centre (TR=1s, 2.5mm isotropic resolution, 260 volumes). Corresponding structural images were segmented in FreeSurfer7.4.1 with 5 ROI sets: 84 ROIs defined following the Desikan-Killiany atlas (DK); a finer parcellation of 360 cortical ROIs (Glasser); one with 14 subcortical DK-ROIs (Glasser-DK); 168 ROIs of the AAL3 parcellation; and 70 subcortical of these added to Glasser's (Glasser-AAL). [4] After preprocessing with CONN22.a, individual ROI-to-ROI connectivity matrices (RRCs) were calculated for all 5 ROI-sets, then thresholded by density. [5] We calculated graph metrics for each network using Brain Connectivity Toolbox (BCT). Centrality metrics and clustering of nodes were weighed by ROI-sizes. Nodes with the highest 15% of a given metric were defined as subject-level hubs or highly clustered regions. Group-level hubs were defined if the overall median metric value was in the top 15% among all ROIs. Ratios of differences between individual and group-level hubs were compared across thresholds between controls and patients. Overlap of group-level and individual hubs was assessed through metrics integrated over density ranges. Nonparametric Mann-Whitney tests (α=0.05) were used for between-group comparisons.

Individual hubs showed large variance for both groups, with 58-75% difference from group-level hubs for all centrality metrics and thresholds. The differences decreased with higher network density for degrees, betweenness and eigenvector centralities in coarser parcellations (DK, AAL). Differences decreased in the control group, but increased in the patient group for strengths, and degrees in finer parcellations (Glasser-based – Fig. 1). Overall, we found statistically significant differences between the strength-based hubs of the patient and control groups at high network densities (>26%) and fine parcellations, see Fig. 2.

Group differences in terms of highly clustered regions substantially decreased with rising network densities, yet significant differences were found between the groups at multiple densities, mostly in finer parcellations. Increasing network density had the most dramatic effect on results with the AAL-set, both for centrality metrics (exception for strength) and clustering. This may be due to large size differences between ROIs in this set, which distorted metric scaling. Based on small-worldness, we identified three threshold ranges: at 1-5% density, networks were sparse, between 6-18% they changed rapidly and around 19% they stabilized. We investigated overlaps between subject and group-level hubs by integrating metrics on the three threshold-ranges. ROIs corresponding to parts of the default-mode network (DMN) were denoted as hubs at all density ranges and ROI-sets. Other consistently identified hubs included inferior parietal areas, superior temporal gyri, frontal and parietal opercular regions. Parts of the middle frontal gyri were also frequently identified as hubs. Medial areas of the DMN, unimodal visual, primary auditory and opercular areas were identified as highly clustered regions. Bilateral parahippocampal regions showed high clustering in Glasser-based parcellations at all densities but only appeared at higher densities in the DK-parcellation.

Our results are compatible with pertinent literature regarding group-level hubs [6], as identified areas that overlap with the DMN and association regions. Interestingly, auditory areas are identified both as hubs and highly clustered regions, a possible effect of scanner noise. Strength-based hubs in networks derived from finer parcellations and higher than 26% density were found useful to identify between-group differences. Confirmation of these results on other patient groups with possible inclusion of other weighted metrics seems beneficial.

Monte Carlo simulations are essential for analyzing and validating microstructural models in Diffusion-Weighted Magnetic Resonance Imaging (DW-MRI), especially when analytical solutions are not available. These simulations require accurate geometric substrates to represent the underlying tissue microstructure [1,2,11] realistically. Over the years, the research community has proposed several substrate generation methods for white matter (e.g., CACTUS, Medusa) and gray matter (e.g., CONCEG), offering varying levels of anatomical detail and biological realism [3–5,9]. In this work, we present CACTUS 2.0, an enhancement of the original CACTUS framework, which was designed for ultra-dense packing of simplified white matter substrates. CACTUS 2.0 extends these capabilities to support hierarchical cellular structures, enabling the modeling of complex, branching obstacles such as neurons. This development retains the high-density packing strengths of the original system, achieving volume fractions of up to 70% in gray matter–like configurations.

This tailored enhancement of the original framework introduces three key features: (1) cellular packing optimization using density-preserving deformation algorithms, (2) hierarchical branching systems with controllable morphology, and (3) integration of pre-segmented neuronal structures. To achieve high cellular densities (up to 95%), we adapted metaball-based techniques with dynamic growth factors that enable efficient packing. For modeling hierarchical structures, we implemented a recursive branching algorithm that supports user-defined control over morphological parameters such as branching angle, segment length, diameter, and distribution. This allows the generation of diverse structural patterns, including pyramidal, targeted, and omnidirectional growth. Lastly, the CACTUS optimization framework was extended to account for complex hierarchical geometries while preserving their morphology during collision resolution and overlap removal.

CACTUS 2.0 successfully generated dense and morphologically realistic cellular substrates. As shown in Figure 1, our growth-based optimization outperforms traditional sphere packing methods, which plateau around 55% density, by reaching up to 95% packing density, producing biologically plausible, non-spherical deformations that align with electron microscopy observations [8]. Figure 2 highlights the versatility of our framework in handling diverse neuronal morphologies. It includes synthetic pyramidal neurons representative of neocortical layers 2–3 (a,b), targeted structures with directional bias (c,d), and isotropic omnidirectional growth (e,f), as well as segmented real neurons (g) from neocortical tissue [6,7]. Figure 3 demonstrates the ability to preserve hierarchical morphology around large obstacles, emulating arrangements such as astrocytic wrapping around somas. Figure 4 shows a full-scale substrate integrating 350 neurons (based on real templates), within a 200 μm × 200 μm × 100 μm volume. After optimization, all overlaps are resolved, and the resulting packing density reaches 70% in the central part of the voxel, a benchmark previously out of reach for neuron-level simulators.

CACTUS 2.0 significantly expands the modeling scope of substrate generation tools by incorporating hierarchical, biologically-inspired structures such as neurons and glia. By supporting synthetic and segmented cellular morphologies, the framework enables the construction of ultra-dense and structurally diverse substrates suitable for simulating gray matter microenvironments. The primary challenge now shifts from achieving dense packing to reproducing biologically realistic spatial arrangements across brain regions. Although CACTUS 2.0 can efficiently integrate microscopy-derived structures, capturing region-specific patterns—including accurate soma placement, branching orientation, and local cellular diversity—requires further anatomical guidance. Solving these new challenges will unlock a new potential for using these substrates in geometric modeling and studies that demand functional fidelity, such as connectomics and DW-MRI signal interpretation.

CACTUS 2.0 represents a substantial advance in computational modeling of brain tissue microstructure. Its capacity to achieve packing densities above 70% while preserving complex, hierarchical morphologies makes it a powerful tool for simulating tissue architectures relevant to diffusion MRI and microstructure-informed analysis. CACTUS 2.0 lays the groundwork for more realistic signal simulations and validation studies by enabling the generation of high-fidelity, voxel-scale substrates.

Mild cognitive impairment (MCI) is a transitional stage between normal aging and dementia, particularly Alzheimer’s disease (AD) [1], marked by cognitive decline, especially in memory, without major disruption to daily life [2]. Diffusion MRI (dMRI) assesses white matter integrity [3], and previous studies using tractography have revealed reduced connectivity in MCI, particularly in memory- and attention-related pathways [4]. However, conventional diffusion tensor imaging (DTI) is limited in areas with free-water (fw), skewing metrics like fractional anisotropy (FA) and mean diffusivity (MD) [5]. Fw correction of multi-shell dMRI improves accuracy by separating tissue and fw signals [6]. However, few studies have examined the impact of fw contributions on tractography. This study compared standard vs. fw-corrected dMRI (tractography and DTI metrics) using Alzheimer's Disease Neuroimaging Initiative (ADNI) data [7] from healthy controls (HCs) and MCI patients. We also assessed how diffusion directions affect tractography and explored structural connectivity differences, linking them to cognitive performance. By combining fw correction and multi-shell dMRI, this work provides deeper insights into white matter changes in MCI.

The study included 153 participants (105 HCs and 48 with MCI). Cognitive function was evaluated using the Mini-Mental State Examination (MMSE) [8], and apolipoprotein E (ApoE) status was assessed as a genetic risk factor for AD [9]. Multi-shell dMRI data were collected on a 3T Siemens scanner, with 127 diffusion directions at b-values of 500, 1000, and 2000 s/mm². Preprocessing steps included restructuring the dMRI data into subsets of 70, 90, and 110 directions to generate synthetic datasets, applying noise reduction, and performing free-water-corrected tractography analysis using a multi-tissue constrained spherical deconvolution approach [10]. Figure 1 summarizes the demographic, cognitive, motion, and APOE characteristics of the groups. Synthetic dMRI datasets with varying gradient directions and noise levels were created to examine the effects of angular resolution and free-water correction on tractography accuracy. Structural connectivity differences between HC and MCI groups were analyzed using both standard and free-water-corrected tractography. Additionally, a region-of-interest (ROI) analysis was conducted on diffusion metrics for connections showing significant group differences. Statistical comparisons included paired t-tests (standard vs. corrected metrics across different directions) and independent t-tests (group differences in connectivity and ROI metrics), with sex, age, and motion as covariates. Effect sizes were computed using Cohen’s d (Figure 2).

The fw model yielded significantly longer fiber lengths (p < 0.001) and lower false positive rates (p < 0.01), whereas the standard model produced shorter fibers and higher false negatives (p < 0.001). Overlap fraction decreased in some cases, particularly for the fw model at 70/90 directions, while orientation error remained similar across models. Figure 3(a) displays standard tractography-derived connectivity differences between groups. The accompanying table reveals HCs showed stronger connectivity overall, except for the right path between Heschl’s gyrus and the right temporal pole, where MCI showed increased connectivity. Figure 3(b) presents the results using fw-corrected tractography, demonstrating uniformly higher connectivity in HCs, especially involving the left middle frontal gyrus (MFG.L). Figure 4 compares standard DTI (a) and fw-corrected DTI (b) metrics between groups. No significant differences emerged with standard DTI, but fw-correction identified a higher fw-index in MCI for the right inferior frontal gyrus (IFGoperc.R) to right insula (INS.R) pathway (t = 3.155, p = 0.002), underscoring its enhanced sensitivity to microstructural changes.

Both standard and fw-corrected tractography revealed reduced structural connectivity in MCI compared HCs. Notably, fw-corrected tractography demonstrated stronger statistical effects than standard methods, suggesting improved sensitivity to group differences. While standard diffusion metrics failed to detect significant changes, the fw-corrected model identified a higher fw-index in MCI for the right inferior frontal gyrus (IFGoperc.R) to right insula (INS.R) pathway. This finding supports the hypothesis that fw correction enhances detection of early microstructural changes in white matter, which may precede overt cognitive decline.

Fw-corrected tractography provides a more precise tool for assessing white matter degeneration in MCI compared to conventional DTI. The elevated fw-index in the IFGoperc.R–INS.R pathway—a region linked to cognitive control—underscores the potential of fw correction to uncover subtle neurodegenerative changes before they manifest clinically.

Spiral trajectories provide a highly efficient strategy for spatial encoding of the object in MRI (see [1], [2]). However, in practice, image reconstruction of spiral scan data is challenging. Higher-order SENSE reconstruction [3] incorporates information about the spatial sensitivity profile of receiver coils, the static off-resonance field (B0) and higher-order gradient field perturbations. It is a generic approach as it renders image reconstruction as a linear inverse problem. However, generating accurate sensitivity maps and especially B0 maps difficult. We present two algorithms to address both issues. Although this type of SENSE reconstruction is computationally intensive, we demonstrate that fast image reconstruction is feasible using an optimized implementation running on modern GPUs.

Single subject data measured using the 16-channel Skope NeuroCam 3T (NeuroCam 3T, Skope Magnetic Resonance Technologies AG, Zurich, Switzerland) on a Philips Achieva scanner (31mT/m gradient strength, 200T/m/s slew rate, Philips Healthcare, Best, The Netherlands) has been used for all analysis/results presented in the following. Both sensitivity and B0 maps are derived from images acquired during the same prescan. The corresponding sequence parameters used are provided in Figures 1 and 2. Following the reconstruction of single-coil images, the first temporal eigenvariate is removed to generate the initial sensitivity maps. This removed eigenvariate, which represents a coil-combined image, is also used to compute the trusted and recon mask. The trusted mask includes voxels with sufficient SNR to reliably estimate sensitivity maps, while the recon mask includes all voxels which might contain the object. The smoothing and extrapolation algorithm is then applied, enforcing consistency with the initial sensitivity maps within the trusted mask and penalizing the second derivative in all directions within the recon mask (see Figure 1). The sensitivity maps are used to compute complex-valued, coil-combined images from the prescan data. An initial B0 map is then generated by performing a linear fit to the phase evolution across different echo times. The smoothing and extrapolation algorithm enforces consistency with the initial B0 map and penalizes the first derivative, in all directions (see Figure 2). The strength of this regularization, or penalty, is made locally proportional to the standard error of the linear fit. This formulation makes the algorithm preserve edges at location where the initial B0 map is accurate. Hence it behaves similar to total variation denoising (see [4]). However, it only requires the solution of a linear system of equations. Within higher-order SENSE reconstruction a large linear system of equations has to be solved using the iterative CG-method [5]. A GPU-optimized implementation is used the reconstruct the 2D spiral scan data (see Figure 3). All allocated variables simultaneously fit into the memory of a commercially available GPU. No recomputation of the encoding matrix during each CG-iteration, as suggested in [3], is necessary. Most operations involving vectors and matrices can be highly parallelized on a GPU which results in a considerable speed-up.

The smoothing and extrapolation of the sensitivity maps and the B0 map is depicted in Figures 1 and 2. Both smoothed and extrapolated maps look realistic and appear to have an increased SNR. We consider the average runtimes of the most demanding operations indicated in Figure 3 as exceptionally fast. Multiple reconstructed images are shown in Figure 4. Image quality is comparable to what is presented in [1].

The smoothing and extrapolation algorithms are capable of either sensitivity and B0 map enhancement, enabling high-performance SENSE reconstruction. Since SENSE reconstruction relies on an explicit signal model of the measured coil data, we suspect the computed maps to be physically meaningful and realistic. Although higher-order SENSE reconstruction is computationally intensive, the GPU-optimized implementation enables fast image computation, effectively alleviating a key limitation of this method. The cost of the used GPU is negligible compared to the cost of an MRI scanner or an NMR field camera. However, for 3D spiral imaging the available memory might not suffice yet. High-quality images were reconstructed even in challenging scenarios. The long acquisition time of 71ms makes the R=2 spiral scan particularly sensitive to local off-resonance effects, while the high undersampling factor of R=6 could introduce aliasing. Nevertheless, these typical artifacts are hardly visible in the reconstructed images.

Using accurate sensitivity and B0 maps high quality single shot spiral images can be reconstructed. SENSE reconstruction can be accelerated using an GPU-optimized implementation. Our work makes spiral imaging using SENSE reconstruction practical.

Diffusion weighted imaging is often a time-consuming procedure as signal averaging is applied for improving the SNR. In a previous work we proposed a novel framework where, instead of signal averaging at very few b-values, 21 linearly spaced b-values over a wider range are measured without averaging [1]. In this context a model-based iterative algorithm was applied to estimate and to correct for Rician bias [2], resulting in more robust ADC maps. However, the proposed framework relies on pixelwise non-linear regression leading to computation times of several hours for a single patient. As this is impractical in a clinical context, parameter estimation by means of deep learning can significantly decrease computation time and might be a more viable option [3,4]. Furthermore, noise reduction in DWI signal through deep learning-based reconstructions have shown potential in improving image quality [5-8]. Expanding on previous works, deep learning reconstructions based on convolutional neural network (CNN) architectures were considered here for denoising prostate DWI. This approach to deep learning-based reconstructions reduces computation time and achieves similar image quality to that of model-based techniques.

Informed consent of the patients was received prior to imaging [9]. A multi-b sequence was used to acquire diffusion-weighted (DW) images at 21 linearly distributed b-values, ranging from 0 to 2000 [s/mm²], for three orthogonal diffusion encoding directions. A self-supervised training was implemented, as depicted in Figure 1, together with a five-fold cross-validation. The acquired DWIs, excluding the b0-image due to perfusion effects, were used both as input and the target for the neural network (NN). The resulting output of the NN are parameter maps depending on the fitting model used for denoising. Here, the kurtosis model S(b) = S0*exp(-bD+b²D²K/6) was considered to describe the signal. Subsequently, Rician bias was added by using a Gaussian noise map derived from OBSIDIAN [2]. The output images were then compared with the input using an L1 loss function. During inference the denoised DWIs from the fitting model were used as the output, i.e. no noise map from OBSIDIAN is necessary. In this study, standard UNet architecture [10], Attention based UNet [11] and Residual Attention UNet [12], were evaluated.

The comparison of DW images from the clinical protocol, OBSIDIAN and the network models is shown in Figure 2. The images appear similar at low b-values but differences are observed at higher b-values, with images from OBSIDIAN and the NNs having lower intensity compared to the clinical images. Some anatomical structure seen in clinical images at b= 100 [s/mm²] vanishes at higher b-values, but to a lesser extent in OBSIDIAN and the different NNs. As seen in Figure 3, the overall appearance of the ADC maps from the NNs are similar to OBSIDIAN with preserved anatomical structures and intensity distribution. ADC maps of the network models and OBSIDIAN, in comparison to raw- and clinically derived ADC maps, appear brighter in all areas as a result of performing bias correction. There were only subtle differences in the overall appearance between the NNs. The estimated ADC values in Figure 4 by the network models are higher than clinical ADC and similar to OBSIDIAN. The ADC estimated by Residual Attention UNet were most comparable to that of OBSIDIAN. Overall, all employed models yielded reasonable estimates with varying degrees of uncertainty and ADC values across different network models.

The visual quality of NN-generated DW images is very similar to the model-based algorithm OBSIDIAN, and visually more appealing compared to clinical DW images because of the lower signal intensity outside the prostate due to Rician bias reduction. With ADC maps no significant improvement in visual quality was observed. However, with ADC estimations both denoising methods result in higher estimates than clinical ADC as a result of bias reduction. In this study DW image data was acquired from 25 patients, with test- and training data from 2 and 23 patients respectively. The size of the dataset constitutes a limitation and a larger dataset can be expected to improve the NN-based denoising. Using a UNet model in denoising for a single patient takes ~10 s with a GeForce GT 1030, which is a significant advantage over model-based algorithms and viable in clinical practice. Before total clinical integration, the method can be improved by including the noise information in the input data, and by using scanner-derived noise maps instead. Training with a larger dataset is also needed for robust performance. In the future, other model functions can be considered and more advanced NN-models, such as vision transformers (ViT), can be explored.

In conclusion, NN-based denoising of DW image data has potential in clinical use and in achieving similar image quality to model-based algorithm with drastically reduced computation time.

MRI offers good soft tissue contrast but long scan times[1].Traditional methods like Parallel Imaging (PI)[4] and Compressed Sensing (CS)[5] have improved reconstruction, but Deep Learning(DL) with Convolutional Neural Networks(CNNs) [8] offers superior results with fewer artifacts.However, DL faces challenges, particularly in applications like dynamic contrast enhancement and cardiac cine, where the extensive acquisition of fully-sampled data for training is time-consuming and impractical[6]. Recent studies have explored GANs for unsupervised MRI reconstruction.UC-GAN[11] allows training with limited data, avoiding the need for fully-sampled ground truth.It combines the Iterative Shrinkage-Thresholding Algorithm(ISTA)[14] with a data consistency step to handle complex data. To improve training stability, variants like WGAN and WGAN-GP add refined loss functions[13].However, UC-GAN and similar models face limitations, including high computational cost, unstable training, and poor perceptual features.To address these, we propose OLP-UC-GAN:a lightweight UC-GAN that uses Depthwise Separable Convolutions(DWSConv) [16] to reduce computation and Perceptual Loss[17] to enhance visual quality. It learns directly from under-sampled k-space data, improving both reconstruction accuracy and efficiency for better clinical utility.

We propose OLP-UC-GAN for MRI reconstruction, integrating DWSConv[16] into the generator of UC-GAN[11] to enable a lightweight architecture.The generator is based on ISTA[14]; and a Perceptual Loss from a pre-trained VGG-19[17].Proposed model discriminator employs a CNN [8] with residual connections and Leaky ReLU activations to ensure stable gradient flow. Proposed model generator receives under-sampled complex k-space data as input and provides the reconstructed 2D complex-valued image.A data consistency step is also applied on the generated image to align the reconstructed image with the acquired k-space measurements by transforming it back to k-space, modulating with coil sensitivity maps, and masking. The discriminator then distinguishes between the real and reconstructed k-space measurements. Training and evaluation are conducted on complex-valued 3T human knee MRI data available online[12], acquired with an 8-channel coil and a 3D FSE sequence. ESPIRiT is used for coil sensitivity estimation[15], and the data is under-sampled retrospectively with a uniform density mask at an acceleration factor of R=4. In the context of MRI reconstruction, the multi-channel Fourier acquisition can be written as (Eq. 1)[11] : x = Bz + ɛ where x is the k-space data, B is the imaging model (including subsampling, Fourier transform, and coil sensitivity modulation), z is the reconstructed image, and ɛ is complex Gaussian noise. B is sampled from a known distribution p_B, leading to k-space data distribution pₓ.The conditional GAN objective is[11] :min┬G⁡ max┬D⁡ E_(x~p_(x ) ) [q (D(x))]+ E_(x~p_(x ),B~p_B ) [ q (1- D(B (G(x))))]( Eq.2)Here, G is the generator, D the discriminator, and q(v) = v denotes the WGAN-GP loss [13], promoting stable training. Additionally, a Perceptual Loss [17] based on VGG-19 features is added to ensure visual similarity between G(x) and the ground truth, improving high-level feature fidelity. The final objective thus combines adversarial and perceptual components to enhance reconstruction quality.

Figure3 illustrates that the proposed method significantly improves knee MR image reconstruction over CS [5] and UC-GAN [11].Table1 presents quantitative results using PSNR, NRMSE, and SSIM.The proposed method achieves the highest PSNR (28.68 vs. 20.89 for UC-GAN and 18.75 for CS), the lowest NRMSE (0.26 vs. 0.38 and 0.75), and the highest SSIM (0.94 vs. 0.88 and 0.68), indicating superior reconstruction accuracy and structural fidelity.

The proposed OLP-UC-GAN model introduces key enhancements that significantly improve MR image reconstruction.Integration of DWSConv[16] into the generator architecture results in better preservation of fine structural details and higher quantitative metrics, while also reducing model parameters for improved computational efficiency, which is critical for enabling faster processing and deployment in resource-constrained or real-time clinical environments.Additionally, incorporating Perceptual Loss[17] into the objective function aligns the reconstructions more closely with high-level features of the ground truth,enhancing visual fidelity.These modifications collectively demonstrate the effectiveness of the proposed method in achieving accurate and efficient reconstruction.

The proposed method combines DWSConv and perceptual loss to enhance both computational efficiency and image quality.Results show improved MRI reconstruction across key metrics, highlighting the value of advanced convolutional techniques in deep learning-based medical imaging.This approach ensures accurate reconstruction while preserving structure and reducing computational cost.

Arterial spin labeling (ASL) is a promising non-invasive MRI technique for perfusion quantification, but it still faces the challenges of low signal-to-noise ratio (SNR) and relatively long scan times. This is especially true for multi-pulse-delay measurements to capture temporal regional variations of the magnetically labeled blood bolus. Typical ASL scans are based on Cartesian data acquisition, where the data generated are assigned to a time point (k0). To obtain time-resolved information, the scan must be repeated at the new time point, resulting in proportionally longer acquisition times. In contrast, radial sampling allows continuous acquisition of perfusion-weighted images (PWIs) with multiple post-labeling delays (PLDs) within a single scan [1]. Expected advantages are higher time efficiency, higher subsampling, and robustness to motion artifacts. Limitations of continuous radial acquisition are the signal attenuation of the labeled blood bolus and the still low contrast-to-noise ratio (CNR). To enhance image quality, particularly at later PLDs, image reconstruction is performed using the Berkeley Advanced Reconstruction Toolbox (BART) [2] with total generalized variation (TGV) [3] and infimal convolution of total generalized variation (ICTGV) [4] regularization, incorporating ASL-specific regularization on the PWIs [5].

A radial FLASH sequence was extended with a pseudo-continuous arterial spin labeling (PCASL) [6, 7] protocol (see Fig. 1). Sampling followed a golden-angle pattern [8], using a 5° flip angle and TR = 2.5 ms. The golden angle was incremented across averages while maintaining the same spoke pattern for each control/label pair. Two-dimensional ASL brain measurements were conducted on a 3T scanner (Siemens Healthineers, Erlangen, Germany) in a healthy volunteer. For continuous measurements, 260 spokes were acquired per average with a labeling duration (LD) of 1800 ms and a post-labeling delay (PLD) of 800 ms. To evaluate the signal quality of labeled blood flow from the continuous measurement, single-delay (SD) images with 65 spokes were acquired as ground truth for comparison. A total of 20 averages were acquired for each measurement. For image reconstruction k-space data was segmented into four frames, with each frame reconstructed from 65 consecutive spokes. Images were reconstructed using the parallel imaging compressed sensing (PICS) [9] tool from BART, applying ASL-TGV or ASL-ICTGV regularization with a spatio-temporal constraint on the PWI [5]. This constraint enables joint emphasis on spatial and temporal features, which is beneficial for the temporal dynamics critical to ASL imaging. PWIs were averaged over 20 repetitions. Reference reconstructions without (ASL-specific) regularization were performed. Additionally, the same k-space dataset was highly undersampled by dividing it into 12 frames, each reconstructed from 21 consecutive spokes.

Fig. 2 compares continuously acquired PWIs reconstructed with different regularization methods (no regularization, TGV, ASL-TGV, ASL-ICTGV) to reference SD images. PWIs at later PLDs appear more detailed and structurally similar to their SD reference when ASL-specific reconstruction is performed. In contrast, without ASL-specific reconstruction, PWIs at later time points become increasingly noisy. Fig. 3 shows the reconstructed PWIs from the highly undersampled k-space dataset. Similar improvements are observed as in Fig. 2, but the effects of strong undersampling further highlight the advantages of ASL-specific regularization. The ASL-TGV and ASL-ICTGV reconstructions yield PWIs of reasonable quality for later time points, whereas the corresponding TGV results are nearly indistinguishable from noise. No visible differences were observed between ASL-TGV and ASL-ICTGV regularization.

The radial readout scheme enables continuous perfusion signal acquisition within a single scan, which may be beneficial for kinetic model fitting to estimate physiological parameters. This approach may also enhance robustness to motion, potentially improving reliability in challenging subjects. The integration of TGV/ICTGV regularization combined with an ASL-specific regularization on the PWI in BART, facilitates the reconstruction of continuously acquired radial ASL data with improved image quality, particularly at later PLDs. These advanced regularization strategies are especially valuable in ASL due to its inherently low SNR.

We successfully combined continuous radial ASL data with ASL-specific TGV and ICTGV regularization, resulting in improved image quality at later post-labeling delays. Future work will focus on extending this method to 3D for whole-brain coverage, enhancing labeling efficiency through background suppression, and comparing its performance to conventional Cartesian ASL sequences.

The Berkeley Advanced Reconstruction Toolbox [1] is an MRI reconstruction framework and recently received support for streaming and real-time reconstruction [2]. Real-time reconstruction is especially useful when directly connected to the MRI scanner, as this enables e.g. live-streaming MRI images from the body, which can improve interventional MRI. In this work, we describe a method for measuring end-to-end latency of real-time MRI systems and apply it to an iterative real-time reconstruction method [3] implemented in BART.

BART contains a set of tools for reconstruction and data/image processing. Typically, several tools are successively applied to the data, usually with a reconstruction script written e.g. in BASH or Python. Previously, each tool would wait until the previous steps were finished. Recently, we added the options to stream data and to let each tool operate iteratively on subsets of the full data [2]. These changes enable real-time reconstruction. Fig. 1 gives a graphical overview of a simple real-time reconstruction pipeline for Cartesian data. We propose an improved [4] method for analyzing the latency of real-time MRI systems, which is described in Fig. 2. Using a long, stiff string (fishing line), a movement is created in the scanner from the control room. This is recorded using a camera both directly by observing a marker attached to the string, and through the MR images shown on the MRI console. A Fairphone 4 (Fairphone B.V., Amsterdam) is used as the camera, which supports a framerate of 120fps. This is considerably higher than the expected real-time MRI framerate and should therefore provide sufficient accuracy for the latency measurements. The video is split into two image series, segmented, and the resulting apparent object speed is calculated. The velocity curves are cross-correlated to obtain the total delay between the marker movement and the movement displayed on the MRI console. This delay represents the end-to-end latency, i.e. the full delay introduced by all components of the real-time MRI system. Measurements were performed on a 3T scanner (Siemens Healthineers, Erlangen), using a radial FLASH sequence with 9 spokes, base resolution 128, TE=1.09ms and TR=1.75ms. The data is streamed to and from the scanner directly in BART streams format, reconstructed using real-time NLINV [3] and subjected to a five-frame median filter.

The latency measurement methodology is validated by assessing the delay between the moving object inside and outside of the scanner, without MRI, i.e. when the camera simultaneously optically observes both ends of the string. We found a delay of 0 frames, confirming that there is no additional offset within measurement accuracy. Fig. 3 shows a snapshot from the video which was processed to determine the latencies. The MR image of the suspended test tube and the marker are clearly visible. Results from the latency measurements are shown in Fig. 4. Initially, there is a high latency of several seconds. The latency decreases and reaches a steady state of approximately 200 ms after less than 3 seconds. This represents an end-to-end latency - the reconstruction time per frame, as determined from timestamps, is much lower (~ 40 ms).

The latency measurement shows that BART reconstruction scripts can achieve good latencies. Time needed for reconstruction is only a fraction of the total latency. Possible reasons for the initially high latency include initialization of GPU memory, starting of processes etc.. One contributing factor to the additional delay observed in the end-to-end latency is the response of the median filter. This explains three frames or about 47.25 ms of the delay. The origin of the remaining latency is currently unclear. Some possible reasons include our custom software for interfacing BART and the scanner, or data processing on the MRI. Concerning the latency measurement methodology, the setup is much simpler compared to our previous method [4].

A simple latency measurement setup was devised and tested, with which we quantified the real-time quality of the BART streaming reconstruction. The time required for the reconstruction step was found to be small compared to the end-to-end latency.

4D flow MRI captures and quantifies blood flow dynamics in three-dimensional space over time. Utilizing k-space segmentation, the data collection over multiple cardiac cycles requires managing substantial 4D data along the spatial and temporal dimensions. Heart rate variability and respiratory motion further complicate the data acquisition process by introducing outliers that can degrade image quality hindering accurate flow visualization. To address these challenges, a range of standard navigator gating and respiration control techniques are utilized [1]. However, these techniques remain demanding and often fail to fully eliminate the motion-induced outliers leading to image degradation. Advanced imaging methods like Compressed Sensing (CS) can also be utilized for 4D Flow MRI reconstruction that allows for sparse and random data acquisition, followed by nonlinear data recovery, maintaining image quality in the reconstruction process [2]. Although CS methods outperform the prior, non-sparsity based techniques, such as SENSE and GRAPPA, they rely on the model [2], which may not hold in case of uncompensated motion. Another approach to handle motion-induced outliers in the collected k-space data is Compressive Sensing with Outlier Rejection (CORe) [3], recently proposed in literature that explicitly models the outliers as additive variables, facilitating the separation of the measured data into outlier estimates and the k-space data for reconstruction [3]. However, this model does not capture outliers efficiently because of the varying types of outliers present in an individual 4D flow dataset due to physiological movements and fluctuating blood. This paper proposes, a new method, which further enhances CORe by differentiating the range and extreme outliers in the k-space data for physiological movements. The efficacy of the proposed method is validated by performing reconstruction on 4D cardiac dataset for the mitigation of motion outliers and was also compared with the conventional CS and CORe reconstructions.

The key steps of proposed methodology are shown in Figure 1. The process starts with 4D cardiac dataset preprocessing through parameter initialization i.e., acceleration rate, number of iterations, regularization parameters etc. as shown in Figure 1. Haar-Wavelet transform was utilized (because of its simple domain operations) to improve reconstruction quality by enforcing sparsity in the wavelet domain. The proposed method is outlined in Equaion (1): x = (1 / σ²) * ‖Ax - y + v₁ + v₂‖₂² + R(x) + λ₂ * ‖g(v₁)‖₁ + λ₃ * ‖g(v₂)‖₁ (1) Where Ax - y is the Fidelity term, R(x) is the Wavelet transform and introduction of variables v₁ and v₂ are used to model the outliers within the dataset range and outside the range. By taking the L₁ norm of the functions g(v₁) and g(v₂) which denote the group of outliers, group sparsity is promoted in the k-space. These sparse outliers are classified into range and extreme outliers by utilizing IQR- Inter Quartile Range method, and a random slice containing outliers is displayed in Figure 2. The outliers are later removed by empirically setting different threshold values during suppression process. For reconstruction, the proposed method employs Split Bregman algorithm [5] as shown in Figure 1. It was chosen for its efficiency in computing sub-problems iteratively by separating data fidelity, sparsity, and outlier’s term, solving them using soft thresholding with Bregman updates. This enables robust and fast convergence of the proposed method. The resulting Four-Dimensional MR images were then reconstructed and post processed to visualize blood flow in the aorta, after background phase correction, adjusting the phase inconsistencies.

Experiments were performed on prospectively acquired 4D flow data of the ascending aorta and is openly available [4]. Figure 2 and Figure 3 highlight the 4D reconstruction results including flow visualization and mean velocity outputs in Aorta, respectively. The proposed method outperforms CS [2] and CORe [3] methods, in terms of PSNR and SSIM values as given in Table 1. While velocity graphs for all the methods might suggest similarities, a more rigorous assessment was performed via Pearson correlation coefficient [6] and R² [7] values between the velocity magnitude image and the ground truth, where proposed method shows better accuracy in terms of reconstructed image magnitude and velocities, thus contributing to a clearer and more interpretable depiction of cardiac flow.

This paper introduces a novel reconstruction method to mitigate the motion outliers in 4D flow MRI and compares its performance with conventional CS and CORe methods. The results from 4D Cardiac dataset show that the proposed method is more effective in suppressing motion outliers while preserving high image quality, making it a superior approach for accurate 4D flow MRI reconstruction.

Magnetic resonance imaging (MRI) is one of the most versatile imaging techniques in modern clinical settings. Besides high-field and ultra-high field applications, MRI has advanced to low-field configurations, enabling imaging of small probes using tabletop systems. Finally, educational purposes benefit from the recent development of fast MRI simulators [1]. The gammaSTAR sequence development framework offers a versatile tool to implement vendor-agnostic MR sequences to control the mentioned variety of MR systems [REF]. However, to date, image reconstruction was still performed using the vendor-specific reconstruction pipelines, which can be a problem for quantitative MRI tasks or the development of imaging biomarkers as this might introduce variabilities into the image quality on different systems. In this work, we introduce gammaSTAR reconstructions which overcome the current limitation by implementing a fully functional reconstruction server, connected to the gammaSTAR framework. With this, we demonstrate vendor-agnostic MR imaging on a 3T Siemens VidaFit system (Siemens Healthineers, Erlangen), on an Ilumr tabletop system (Resonint, Wellington) and using the MRZero simulation tool.

All gammaSTAR sequences utilize the ISMRMRD standard [3] to assign relevant measurement flags to individual data readout events. This allows to implement a highly flexible reconstruction server, which applies the same series of reconstruction modules for all non-Cartesian (using mri-nufft toolbox [4]) and Cartesian acquisitions (cf. Fig 1a). Whether a module needs to be applied is decided based on the received data dimensions and ISMRMRD flags. The reconstruction server is provided as an OpenRecon compatible Docker image. The complete workflow of measuring/simulating data is demonstrated in Fig. 1b. MR sequences can be implemented in a modular fashion using the gammaSTAR sequence frontend. From the frontend, the corresponding sequence protocol can also be adjusted. The adjusted sequence is then sent to a variety of systems using specific drivers. Finally, resulting image data is displayed in the frontend. To validate the flexibility of the reconstruction tool, a large set of different two- and three dimensional sequences is investigated. Using MRZero 2D RARE, 2D RADIAL, 2D FLASH, 2D bSSFP and 2D Spin-Echo EPI sequences were simulated. On the Siemens VidaFit 3T system, 2D RARE, 3D MP-RAGE and 2D EPI sequences were applied. Finally, 2D RARE, 2D FLASH and 2D Spin-Echo EPI sequences were tested on the Resonint Ilumr system. Protocols were adjusted to cover a large range of available options such as golden-angle schemes for the radial acquisition, parallel imaging using external and integrated reference scans as well as partial Fourier setups.

Fig. 2 shows results from simulations with the MRZero simulator and a large variety of contrasts and acquisition strategies. Fig. 3 shows results from the Siemens VidaFit 3T full body MR system using parallel imaging strategies. Fig. 4 shows finally shows results from the Resonint Ilumr system.

MR images were successfully reconstructed using gammaSTAR reconstructions for the Siemens VidaFit 3T system, the Resonint Ilumr system and the MRZero simulator. Applied sequences and protocols covered a large range of possible configurations. This demonstrates the flexibility of the reconstruction server in combination with the gammaSTAR sequence development framework, offering the user the option to test sequences using the MRZero simulator before executing them on real full body or tabletop MR systems. Our solution especially benefits from the close link between the sequence development and respective reconstruction tools. This represents an advantage over other open source solutions such as Gadgetron [4, 5] which aim to support a broad range of sequence frameworks, making it more difficult to maintain a common standard. The reduced quality of partial Fourier reconstructions (cf. Fig. 3) as well as the low signal-to-noise ratio on tabletop systems (cf. Fig. 4) will be improved in the future. In addition, image normalization using a body coil reference will be included. Finally, the reconstruction server will be extended by image processing steps, such as automatic quality assessment (AQUA) and quantification options such as generation of T1/T2 maps or calculation of cerebral blood flow maps in case of Arterial Spin Labeling experiments.

We demonstrated the vendor-independent gammaSTAR reconstruction server in combination with the gammaSTAR sequence development tool. The flexibility of the reconstruction server was verified in three different scenarios: A full body Siemens VidaFit 3T system, a low-field Resonint Ilumr tabletop system and the virtual MRZero simulator. Future work will further extend the compatibility to various other systems. In addition, a fully virtual MR workspace will be investigated which allows direct transition of developed sequences and reconstruction techniques to real hardware.

4D flow MRI is an advanced imaging technique for visualizing blood flow dynamics. It plays a critical role in the early and accurate diagnosis of cardiovascular conditions such as congenital heart disease (CHD), aortic wall disease, valvular heart disease, pulmonary hypertension, and peripheral artery disease. This technique involves acquiring one reference image and three velocity-encoded images along orthogonal directions across multiple cardiac cycles [1]. Consequently, the k-space data spans six dimensions: frequency encoding, phase encoding, slice selection, coil channels, temporal frames, and velocity direction. Such high-dimensional data acquisition results in computationally intensive and time-consuming image reconstruction, which poses a barrier to its widespread clinical adoption. To address this challenge, this study presents a novel FPGA-based hardware accelerator architecture for compressed sensing reconstruction of 4D Flow MRI data. The proposed system leverages the parallel processing capabilities of the Xilinx Zynq-7000 SoC to significantly reduce reconstruction time without compromising image quality.

The architecture, illustrated in Fig. 1a, integrates multiple FPGA based parallel computational blocks to accelerate large-scale complex matrix-matrix multiplications. In the proposed accelerator design, the DMA (Direct Memory Access) facilitates high-throughput data transfer between the accelerator and DDR3 memory via the AXI-Stream interface, minimizing CPU intervention and significantly reducing processor overhead [2]. CPU (ARM core of zynq SoC) functions as memory controller and host for arbitration module in its program memory [3]. The arbitrator is responsible for partitioning data arrays coordinating DMA transfer to enable efficient communication with parallel accelerators, ensuring fast and reliable data movement while maintaining consistency with original dataset. Fig. 1b illustrates the design flow of hardware accelerator to compute the wavelet transform, which serves as a sparsifying basis in the regularization term of the compressed sensing objective [4] and is involved in the gradient computation during reconstruction [5]. As shown in Fig. 2, gradient computation dominates the overall reconstruction process, accounting for approximately 72% of the total reconstruction time High-level-synthesis (HLS) is adapted for designing hardware accelerator, targeting wavelet transform module used in compressed sensing. The algorithm, implemented in C/C++ is optimized using HLS compiler Directives i.e. HLS pipeline, HLS unroll to exploit fine-grained parallelism and maximizing throughput [6]. Design constraints are enforced to meet latency, and area targets suitable for the Xilinx Zynq-7000 platform, achieving a balance between performance and hardware resource utilization. Functional correctness is validated through C/RTL simulation and co-simulation, ensuring bit-accurate correspondence between the high-level model and synthesized RTL.

Reconstruction time comparison between the proposed method and CPU-based implementation is presented in Fig. 3a. The results show that the proposed method outperforms the CPU-based compressed sensing reconstruction, achieving a fourfold acceleration in overall reconstruction time while maintaining image quality. Fig. 3b presents a comparison of 4D Flow MRI reconstructed images (magnitude and three velocity-encoding directions) for a single frame using both the CPU and the proposed method. The results demonstrate that the proposed method achieves image quality comparable to that of the CPU-based reconstruction.

This research introduces a novel FPGA based SoC architecture for accelerating wavelet transform in sparsity gradient of compressed sensing for 4D flow MRI reconstruction. The results using 30-coils, 20-frames in-vivo 4D stress cardiac dataset shows that proposed method can achieve four-times faster reconstruction without compromising the reconstruction accuracy and making 4D flow MRI suitable for clinical adaptation.

Thermal ablation is widely used as a first-line treatment for malignant tumors [1–3], relying on extreme temperatures to destroy tumor tissue. Magnetic Resonance Imaging (MRI) thermometry enables real-time monitoring of temperature changes during procedures [4–6]. However, its precision is limited by measurement uncertainty, reduced accuracy near the applicator due to susceptibility artefact and partial volume effect [7,8]. To address these limitations, numerical heat transfer simulations can complement MRI data by improving spatial resolution and reconstructing temperature information lost due to artifacts or low signal quality [9–12].

Model description: A hybrid framework was developed to simulate temperature distributions during MRI-guided thermal ablation. The model relies on the solution of the heat equation in spherical coordinates, solved using the Laplace transform and finite difference methods. For each energy, a source term was specified as follows: For Laser-Induced Thermal Therapy (LITT), heat deposition is modelled as: Q(r,t)=Q0(t)exp(−βr). For Microwave Ablation (MWA): Q(r,t)=Q0(t)exp(−β√r). For Radiofrequency Ablation (RFA): (Q(r,t)=Q0(t)H(r−r0)), where H is the Heaviside function. A spherical temperature distribution is computed and mapped onto a 1-mm grid. This half-sphere is then extended into a cylinder to represent applicators with axial symmetry. The resulting 3D structure is aligned with MRI images using rigid registration. This workflow is illustrated in Figure 1. These solutions are then mapped onto a 3D voxel grid by averaging the temperature inside each voxel, aligned with the experimental MRI data, and a mask is applied to keep only reliable voxels. The model then estimates the heat source parameters Q0(t), β, and r0 by minimizing the voxel-wise error between simulated and measured temperatures over the entire ablation time. The model focuses on estimating heat source parameters while assuming constant thermal properties, simplifying the calculations without losing accuracy. Lastly a super-resolved, three-dimensional temperature field at a finer resolution than the MRI voxel size is reconstructed. Device and experiments: MWA was performed in bovine liver using a 14G antenna (AveCure, MedWave, USA) under MRI guidance. Ablation lasted 7 min 30 s, targeting 60 °C. A 1.5 T scanner (Magnetom Sola Fit, Siemens) acquired 13 slices every 4 s with a multi-shot EPI sequence (TE = 19 ms, TR = 47.3 ms, FA = 90°, matrix = 128×128, slice thickness = 3 mm, gap = 1.5 mm, FOV = 300×300 mm², BW = 953 Hz/pixel). LITT was tested in 3% agar gel using a 976 nm diode laser (8 W, 1 min) coupled to a 240 μm diffuser (ALPhANOV, France). MRI (1.5 T Avanto, Siemens) monitored the ablation with a single-shot EPI sequence (8 slices, every 2 s; TE = 21 ms, TR = 2000 ms, FA = 40°, matrix = 128×128, thickness = 3 mm, FOV = 158×158 mm², BW = 1445 Hz/pixel). RFA used an MR-compatible catheter (Vision-MR, Imricor) with two microcoils for tracking. Power levels of 5, 10, 15, and 20 W were applied in agar gel for 30 s. MRI thermometry (1.5 T Aera, Siemens) acquired 8 slices every 1 s with a multi-shot EPI (TE = 28 ms, TR = 930 ms, FA = 60°, matrix = 148×148, thickness = 3 mm, gap = 0.3 mm, FOV = 170×170 mm², BW = 1165 Hz/pixel). Acquisition: For all acquisitions, temperature maps were computed using the Proton Resonance Frequency Shift after B0 drift correction.

For each case, simulated and experimental temperature maps were compared, and the model successfully predicted temperature distributions that matched the experimental data. Each simulation step took less than 45 ms, and data alignment with MRI measurements averaged 87 s. Figures 2, 3, and 4 present temperature comparisons for MWA, LITT, and RFA, using a 5×5 voxel kernel centered near the applicator. The normalized root mean squared error reached up to 20 °C for MWA, 3 °C for LITT, and 7 °C for RFA. Higher errors in MWA are attributed to complex near-field effects not captured by the simplified model. In contrast, the compact symmetry of the laser diffuser in LITT yielded highly accurate predictions.

This framework combines analytical modeling, numerical methods, and MRI thermometry to reconstruct thermal fields with improved resolution. By adapting the source term to each ablation technique, the model remains flexible and computationally efficient ( under 2 min total runtime). It recovers temperature in regions affected by artifacts and can predict the maximum temperature at the vicinity of the device without partial volume effect. Future work will integrate perfusion effects and apply the model to in vivo datasets for clinical validation.

Parallel transmission (pTx) is essential in ultra-high field MRI to improve signal and contrast uniformity [1], but requires accurate measurement of the channel transmission profiles (B₁⁺) [2]. The interferometric pre-saturated turbo-FLASH (SatTFL) sequence [3-5] offers a valuable technique for this purpose, but its reliability deteriorates with a large B₁⁺ dynamic range of the combine, typically, circularly polarized (CP), mode [6,7]. In this case, a conventional B₁⁺ reconstruction method can drastically amplify noise propagation and impair subsequent pulse design. This problem was addressed recently in [8] by tailoring the interferometric scheme to the RF coil. However, this may appear insufficient if the dynamic range (max/min) of the interferometric modes (IMs) remains too high. In this work, we consolidate the SatTFL method by introducing a least-squares-based (LS) B₁⁺ reconstruction offering a flexibility in interferometric scheme and minimizing noise propagation.

The satTFL signal [9] depends on the saturation (α) and readout (β) flip angles (FAs), determined by complex Nch-dimensional B₁⁺ profiles (Nch is the channel number) and the corresponding IM matrices A and B (Eq. 1, Fig. 1A). Estimating B₁⁺ from measured noisy data s₁, s₂, … represents an inverse problem, with A and B given, and B₁⁻ and m₀ as additional unknowns. The standard B₁⁺ reconstruction uses the S-std scheme (Fig. 1B), with A and B of the form [X;0;0] and [X;X;Y], where X is a single-column and Y is a square "interferometric" matrix [8,10]. With S-std, B₁⁺ is typically derived using the noise-agnostic workflow M-std (Fig. 1E), which involves computing s₁/s₂, s₃/s₂, … (steps 1-3) followed by solving an “interferometric” linear system (step 4). In the presence of noise, this method can be locally instable for voxels with low s₂ signal or poorly conditioned Y. To address this problem, we propose (i) a method (M-LS) to estimate B₁⁺ via least-squares fitting of the SatTFL signal to Eq. 1, and (ii) a method for optimizing the IMs. (i) We factorize B₁⁺ as B₁⁺=σb₁⁺ with σ=||B₁⁺||, and use a two-step workflow (Fig. 1F): 1) estimate the phasor b₁⁺ from the subset of signals acquired with α=0 (a linear least-squares problem); 2) reinject b₁⁺ into the full signal model and estimate σ by gradient descent (a 1D nonlinear least squares). (ii) M-LS is not limited to the S-std configuration but applies to the more general S-Nsat scheme shown in Fig. 1C, where Nsat images are acquired with α≠0. A and B then take the form [X;0] and [X;Y], where X (resp. Y) is the encoding matrix for the Nsat (resp. Nch) unitary IMs. Heuristically, the X and Y blocks are designed separately, based on a representative B₁⁺ distribution and the objective function in Fig. 1D, Eq. 2. This approach ensures that, for each voxel, at least one IM yields a FA close to the target. The condition number constraint guarantees that the inversion of Y is noise-robust. We applied the proposed approach on a Magnetom 7T MRI using the 8Tx32Rx Nova head coil and a home-made 8Tx32Rx Avanti2 prototype [11], which has a markedly different geometry and a larger CP-mode dynamic range. EM simulations on a spherical agar phantom (SP) returned CVs of 0.29 and 0.45 for the CP mode of the Nova and Avanti2, respectively. We computed S-std, S-2sat, and S-3sat based on simulated SP’s B₁⁺ profiles, scaling the IMs to obtain 〈α〉=60° and 〈β〉=4°. SatTFL signals in the SP were simulated according to Eq. 1 (Fig. 1A) using a complex Gaussian noise source n with std(n)/m₀=1e-3; then, M-LS and M-std (when appropriate) were applied. We repeated this procedure 100 times to derive the average relative B₁⁺ error map, ∆. For Avanti2, we additionally found the optimal 〈α〉 and 〈β〉 for S-3sat via exhaustive search, minimizing the average error 〈Δ〉. The IM schemes were then tested in simulation for a human head model (HM) [12]; the FA-optimized S-3sat was finally applied in vivo on one subject.

The simulation results for Nova and Avanti2 are shown in Figs. 2 and 3. The optimal average FAs for S-3sat were <α>=62° and <β>=8°. In vivo B₁⁺ acquisition (Fig. 4) with Avanti2 took 2 min 20 s, with the M-LS reconstruction of ~30s.

The proposed M-LS B₁⁺ reconstruction algorithm reduces noise propagation and supports a wider variety of schemes compared to M-std. For Avanti2, M-std applied to S-std is prone to large errors (〈Δ〉 = 6.7%). Using M-LS reduces 〈Δ〉 to 3.1%, but the S-3sat scheme is necessary to bring the error down to a level comparable to Nova (〈Δ〉=2.3%). Optimized FAs <α>=62° and <β>=8° further decrease 〈Δ〉 to 1.2%. The SP-based scheme remains applicable for the head (〈Δ〉 = 1.3% on HM), but a more optimized scheme could be established iteratively from a database of in vivo B₁⁺ maps.

The proposed method improves the quality of the reconstructed B₁⁺ data required for pTx pulse design and can be applied to a wide range of transmit arrays architectures. It may also be valuable at higher magnetic fields and for body 7T MRI.

Whole-body (WB) MRI, including diffusion-weighted imaging (DWI), is recommended in patients with metastatic prostate cancer [1]. WB-DWI is used for tumour detection and treatment response assessment [2-4], but long acquisition times limit patient comfort and compromise scanner throughput. Deep learning (DL)-based reconstructions have the potential to aid the acceleration of MR imaging while preserving diagnostic image quality [5,6]. Prior studies show improved image quality in DWI with DL-based reconstruction (DL DWI) but lack cancer type-specific quantitative evaluations [7-14]. Our study aims to evaluate image quality and apparent diffusion coefficient (ADC) estimates using an accelerated DL WB-DWI acquisition compared with standard WB-DWI in patients with metastatic prostate cancer.

Two prostate cancer patient cohorts, scheduled for a clinical WB-MRI examination at any treatment stage, were scanned on a 1.5T MR scanner after verbal consent. Patients underwent routine clinical WB-MRI of 5-stations, and research sequences were performed at the end of the examination. Accelerated DL-DWI was adapted from the routine WB-DWI protocol at our centre, using fewer signal averages and higher parallel imaging acceleration (Table 1). DL-DWI data were reconstructed inline using a prototype deep learning-based variational network [15] trained on healthy volunteer DWI data, alternating between data consistency and learned regularization steps as presented in [11]. Cohort 1 (n = 10, mean age 70 ± 7 years) compared image quality between standard WB-DWI and accelerated DL WB-DWI acquisitions across a 4-station (neck to thighs) WB-DWI acquisition. Two experienced radiologists blinded to the type of DWI sequence, scored image quality for b900, b900 maximum intensity projection (MIP) and ADC images on a 4-point Likert scale (described in Figure 1). Cohort 2 (n = 20, mean age 75 ± 9 years) compared ADC values from both protocols on a radiologically active focal bone lesion (size ≥ 1cm) from a single station. Focal lesions were delineated on standard b900 images, and the volumes of interest (VOIs) were transferred to the corresponding ADC maps for each patient (IDL ADEPT, ICR, UK). The median and interquartile range (IQR) of ADC estimate from all fitted voxels in the VOI for each patient were measured. Statistical comparisons used Wilcoxon signed-rank tests with Bonferroni correction to compare image quality scores, and Bland-Altman analysis to compare ADC estimates.

Cohort 1 The qualitative radiological scoring did not find significant differences between the standard WB-DWI and the accelerated DL WB-DWI protocols (p > 0.0025), with most of the examinations rated as ‘good’ in any of the qualitative metrics considered for both protocols (Figure 1). Exemplar b900, ADC and b900 MIP images acquired with standard WB-DWI and DL WB-DWI are shown in figure 2. Cohort 2 The cohort average lesion volume was 16.7mm3(range: 2.4 – 95mm3). The cohort mean and standard deviation of the lesion ADCs were 121.9 ± 41.5 (*10-5 mm2/s) for the standard WB-DWI and 122.3 ± 43.8 (*10–5 mm2/s) for the accelerated DL WB-MRI. Bland-Altman analysis (Figure 3) showed no significant difference in ADC values between the standard WB-DWI and the accelerated DL WB-DWI protocols (p = 0.93). Similarly, the IQR of the ADC estimates was not significantly different between the two protocols (p = 0.63).

The qualitative assessments demonstrated that an accelerated DWI protocol with DL-based reconstruction can reduce acquisition time by 7min and 10s (or 37%) for a 5-station WB-DWI examination without reducing image quality. This result aligns with previous studies reporting maintained or superior image quality and SNR outcomes in high b-value and ADC images of DL-reconstructed DWI compared to a standard DWI protocol [7-9,11,14]. In our 20-patient cohort, accelerated DL DWI did not affect ADC quantification in bone lesions, consistent with prior studies on WB and spine imaging [8,14]. However, conflicting results have been reported for the breast and liver, with some studies finding no differences [10,12] or significantly higher ADC values [11] for DL DWI compared with standard DWI protocols. Unlike other reports [14], the DL DWI sequence was mainly accelerated by reducing b-value averages, keeping the diffusion mode and scheme unchanged. This minimizes added eddy-current distortions, enables direct protocol comparison, and clarifies DL technology’s impact on ADC quantification A limitation of this study is the relatively small cohort size and single-centre design, which may limit the generalisability of our findings. However, the study included patients with metastatic prostate cancer undergoing WB-MRI examinations at any treatment stage and is therefore representative of this WB-MRI patient population.

An accelerated DWI protocol with DL-based reconstruction reduced the acquisition time of WB-DWI by 37% without compromising image quality or ADC estimates.

Several medical protocols benefit from the wide variety of soft tissue contrast that can be obtained from MRI with high spatial resolution. Nevertheless, acquiring high-resolution 3D MRI volumes for different contrasts can require significant acquisition time. In this respect, multi-contrast reconstruction strategies offer a potential solution [1] by taking advantage of the redundancy of information present in the measurements. Techniques such as compressed sensing and super-resolution (SR) use the anatomical information from one specific contrast to reconstruct another under-sampled contrast [2]. In this work, we study the necessary and sufficient conditions for applying SR to multi-contrast reconstruction from a variational point of view. Quantitative and qualitative results on in vivo T1 and T2 weighted 3D TSE data at different stages are presented.

The 3D SR problem we consider is to reconstruct an isotropic volume X from an under-sampled one Y, so that Y~HX, (1) Where H is the forward operator that establishes the correspondence between the high-resolution object X and the low-resolution acquisition Y. It considers the geometric transformations, down-sampling and the PSF. In addition, if X_Ω represents a reference isotropic volume, the inverse problem (1) can be posed as follows: X ̃"\∈\ " 〖"Argmin" 〗_X 1/2 ‖Y-HX‖_2^2+βR(X,X_Ω ) (2) Where R is a regularization term that contains the a priori information about the volume X ̃ to be reconstructed. In this work, We consider: R(X,X_Ω )=‖DX- e_Ω DX‖_2^2 with 〖 e〗_Ω= 〖DX_Ω〗^2/(〖DX_Ω〗^2+ τ^2 ) (3) Where D is the finite difference operator, e_Ω is an edge map obtained from X_Ω (fixed during optimization), and τ>0 a parameter which thresholds〖 DX〗_Ω. Gradients in (3) share the edge localization and direction [3], then the idea of e_Ω is to extract these features. Accordingly, the term (3) can be interpreted as an edge-weighted Tikhonov regularization. Given that e_Ω∈ [0,1] the smoothing in (3) is applied proportionally to the relative importance of a contour. Experiments: First, reference isotropic volumes in vivo (1mm^3) were acquired using a 3D TSE sequence on a 3T MRI system (MAGNETOM Vida, Siemens Healthineers, Germany) with (TE, TR) = (11, 700) MS and (400, 3200) MS for T1-w and T2-w images. Then, from isotropic volumes, we simulated a down sampled anisotropic volume by using the direct operator H (1 x 1 x 4 mm^3). Then, an isotropic volume X ̃ is reconstructed by optimizing (2). Iterative methods were implemented to solve (2), with a relative cost of 10^(-7) order. Regularization parameter β was tuned to maximize the Peak-Signal to Noise Radio (PSNR) and the Structural Similarity Index Measure (SSIM) see Table 1.

In order to study the influence of structure knowledge in reconstructions, four cases were considered: (1) Ground truth contour map: Information about the same contrast is used, that is, if a T1-w volume is reconstructed, then 〖DX〗_(Ω_T1 )is used in (3). (2) A mutual contour map: 〖 e〗_(Ω_(T1∩T2) )= (〖|DX〗_(Ω_T1 ) | 〖|DX〗_(Ω_T2 ) |)/(〖|DX〗_(Ω_T1 ) | 〖|DX〗_(Ω_T2 ) |+ τ^2 ) (3) Constraint contour map:〖 e〗_Ω contains only information from the other contrast, that is, if a T1-w volume is reconstructed, then 〖DX〗_(Ω_T2 )is used in (3). And (4) A gradient projection P_(e_Ω ) proposed in [2] (〖 e〗_Ω contains only information from the other contrast) where R(X,X_Ω )=‖DX- e_Ω ⟨e_Ω│DX⟩‖_2^2 with 〖 e〗_Ω= (DX_Ω)/√(‖DX_Ω ‖^2+ τ^2 ) (4)

A global view for T1-w and T2-w images of an axial slice is showed in figure 1. Figure 2 shows a zoom of a coronal slice. In general, significant and consistent high-resolution content is recovered in HR images (blue arrows), as well as vertical artefacts due to sagittal under sampling (red arrows) are corrected. When a ground truth contour map is used fine details are recovered which is reflected in quantitative measurements (see Table 1). In a second stage, a mutual contour map test showed to be very important, since we found that informations about edge localization and direction are presents in the other contrast. However, when only information from other contrast contour map is considered, unrealistic intensity jumps artefacts are produced (yellow arrows). Hence, a complementary a priori about gradient information must be made, for example the gradient projection proposed in [2] allows for each gradient direction to include the information from the other directions, this attenuates the intensity jumps artefact without losing the gained structures.

The experiments conducted in this study allow us to verify and demonstrate that a priori information about edge location, direction and weighting are the necessary and sufficient conditions for satisfactorily solve the SR problem. This work suggests that multi-contrast SR is feasible if a high-resolution contour map is available. Reconstruction methods that inherently provide edge information such as DMS [4], will be investigated in this regard.