At 3 T MRI, the radiofrequency (RF) wavelength becomes comparable to body dimensions, causing dielectric artifacts and dark regions in images. Recent work suggests using metamaterials [1,2] and metasurfaces (MSs) [3] to shim the transmit field passively. Placed on the body, these structures localize the RF field within the ROI, increasing |B₁⁺| amplitude and mitigating dielectric artifacts. Here, we propose an MS comprising six concentric split-ring resonators (SRRs). This configuration offers two key advantages: easier tuning using four capacitors per SRR and adjustable RF field distribution through frequency detuning (Δf) between SRRs. By varying Δf between adjacent rings, the system achieves tailored B₁⁺ field distributions optimized for specific anatomical ROIs.

Numerical simulations were performed in CST Studio Suite 2022 using a frequency domain solver. The MS consisted of six square concentric SRRs with side lengths of 46, 90, 134, 178, 222, and 266 mm, respectively. Each SRR was made of 2 mm-wide copper strips and included four gaps loaded with variable capacitors (Fig.1a). The resonant frequency of the entire device was tuned to 123 MHz. A detuning approach was applied by selecting the capacitance values such that each inner SRR had a first resonant mode at a higher frequency than its neighboring outer ring. Several configurations with different frequency gaps (Δf) between adjacent SRRs were studied (Fig.1b). To estimate the |B₁⁺| and the SAR distributions, we performed numerical simulations using a voxelized human model ‘Duke’ (Sim4Life) with realistic tissue properties. The transmit RF coil was a 16-leg shielded high-pass whole-body birdcage coil with a 700 mm bore diameter (Fig.1c). The |B₁⁺| and SAR distributions were normalized to 1W of the accepted power. The coefficient of variation (Cv) was used to assess |B₁⁺| inhomogeneity. It was calculated as the standard deviation divided by the mean |B₁⁺| in the ROI, expressed as a percentage. The local maximum SAR (SARav.10g) was calculated as averaged over 10g of tissue. SAR efficiency was determined as the |B₁⁺|/√SARav.10g. Experimental validation was carried out on a 3 T Siemens Magnetom Trio system using a birdcage coil for transmission and a matrix surface coil for reception. The MS configuration with Δf = 7 MHz was selected for experimental studies based on its optimal SAR efficiency in simulations. The HASTE pulse sequence (TE/TR = 61/2000 ms; FA = 180°; FOV = 400×400 mm²; acquisition matrix = 384×384; slice thickness = 5 mm) was used to acquire abdominal MR images of a healthy volunteer with and without the MS. The SNR maps were calculated by dividing the mean signal intensity in the liver region by the standard deviation of the noise estimated in signal-free background areas.

Numerical simulations showed that the MS increased the |B₁⁺| amplitude in the liver from 0.15 to 0.19–0.20 µT (~35% gain). Cv of the B₁⁺ field in the liver was 14.7% in the reference case and changed to 21%, 17.5%, and 19% for Δf = 0, 7, and 20 MHz, respectively. SAR maps (Fig. 3) showed localized hotspots near the liver in all MS configurations. The SARav.10g maximum increased from 0.14 W/kg (reference case) to 0.23, 0.20, and 0.21 W/kg for Δf=0, 7, and 20 MHz, respectively. However, the reference case showed SAR efficiency of 0.40 μT/√W/kg, while the MS configurations demonstrated SAR efficiencies of 0.40, 0.45, and 0.43 μT/√W/kg for Δf=0, 7, and 20 MHz, respectively. Experimental studies of liver MRI demonstrated that using the MS increased the average SNR in the ROI by 46% compared to the case without the device (Fig. 4).

Numerical simulations revealed that all considered MS configurations effectively reshaped the B₁⁺ field of the birdcage coil, producing an increase in B₁⁺ field amplitude within the ROI. This redistribution successfully compensated for the B₁⁺ field minimum present without the MS. Although each configuration increased the local SARav.10g, the concurrent enhancement of the B₁⁺ field amplitude led to improved SAR efficiency. The configuration with Δf = 7 MHz demonstrated the highest SAR efficiency of 0.45 μT/√W/kg, making it the optimal choice for experimental implementation. Experimental MRI studies with a healthy volunteer demonstrated that the proposed MS enhanced signal intensity in the ROI. The current study focused on optimizing the MS's configuration specifically for liver imaging. However, the proposed tunable MS offers broader applicability due to its flexible adaptability to different ROIs. By modifying capacitor values, the spatial distribution of the RF magnetic field can be precisely controlled, allowing customization for different anatomical regions. The tunability and the structure's relative simplicity make the design versatile and practical for clinical implementation. This study was supported by the state assignment No. FSER-2025-0018 within the framework of the national project “Science and Universities,” Russian Federation.

Although MRI is one of the most valuable and versatile imaging modalities, it also has considerably higher operational costs in comparison to for example CT or ultrasound. The relatively long duration of MRI investigation and the need for highly trained staff are the dominant contributors to the operational costs of a MRI system. Reducing setup time and simplifying the setup procedure may help to reduce MRI costs. One of the ways to simplify the patient setup is by removing the need to connect RF coils. Approaches that use integrated circuits to digitize the MRI signal on the coil and use conventional wireless signal transfer have been proposed, but come with i.e. lack of required bandwidth, complex circuitry and full redesign of the MRI receiver chain. A simpler approach is to use the RF body coil that through mutual inductance with a local receiver can obtain the signals from the local resonant coil via the receiver connected to the RF body coil [1,2]. However, this hinders the use of wireless local receiver arrays. Inductive matching has been proposed and implemented for more than 30 years [3,4] in fixed mechanics between the receiver coil and the matching coil. Last year we demonstrated that such mechanics can be made variable while maintaining good matching conditions between the receiver coil and the inductively coupled matching loop [5]. In this work, we investigated if an array of stretched coils that have an inductive coupling with loops made inside the bed can be used as a simple approach for wireless RF coil array operation. More specifically, the feasibility of wireless reception using a 4-channel wireless coil with 4 independent receiver loops was investigated.

A 4-channel wireless coil (6x47cm each channel), (Fig.1.A) was tuned and matched at 64MHz for a 1.5T system (Ingenia, Philips, Best, The Netherlands) when it was loaded with a body phantom (σ=0.55S/m, ε=74 at 100MHz). Decoupling via overlapping was implemented. A passive detuning (BAV99W, Eindhoven, The Netherlands) was applied (Fig.1.B) to decouple each channel from the transmit coil during RF-transmit. Two pairs of 5 cm diameter receiver loop coils were designed (Fig.1.C). Each one was tuned and matched at 64MHz when it was placed 2 cm away from the loaded wireless coil (mimicking the patient table). A detuning circuit was applied (Fig 1.D). Overlapping was implemented to reduce coupling between loop coils. The S-parameters of the receiver loop coils and the 4-channel wireless coil, including efficiency and decoupling, were measured as in Setup1 (Fig.2.A) and Setup2 (Fig.2.B). Phantom Imaging The wireless coil was placed around the phantom. Receiver loop coils were placed 2 cm above the wireless coil and only loop coils were connected to the interface box (Fig.2.C). 2D-Dynamic-GRE sequence was acquired. SNR maps per channel and noise correlation matrix were calculated.

All VNA measurements are given in Fig.3. S21 of Setup1 when the receiver loop coil was placed 2 cm above the wireless coil was -27 dB for all channels (Fig.3.A). The decoupling between channels of the wireless coil of Setup2 was at least -13 dB (Fig.3.B). S11 of each receiver loop coil was less than -14 dB, the decoupling between receiver loop coils was at least -14 dB (Fig.3.C). SNR maps obtained from each channel of the wireless coil, particularly in the coronal plane, show signal intensity peaks at the edges corresponding to the sensitivity profiles of both receiver loop coils and wireless coils (Fig.4.A). The noise correlation matrix shows low noise correlation between the channels (Fig.4.B), resulting in largely independent behavior of the receiver loop coils, which is also in agreement with bench measurements.

This work presents an implementation of wireless MR signal reception via inductive coupling for a 4-channel wireless coil. The bench results of Setup1 and Setup2 show the applicability of the wireless array coil design when all channels are independent. SNR maps and noise correlation matrix show the possibility of independent reception of each channel of a 4-channel wireless coil in phantom studies. It provides a perspective for further steps in the design of a wireless coil array and its accompanying receiver loop coil array that would accelerate and simplify studies of complex regions of the human body.

In conclusion, this study has shown that all 4 wireless channels can operate independently, allowing true multi-channel reception with high sensitivity to be achieved over a wide area. The results presented may provide an impetus for the development of wearable coil array technology.

The first in-vivo human brain images at 11.7 T were recently shared to the community [1]. Such a high magnetic field comes with several challenges, in particular regarding B0-related artifacts in fast imaging sequences like EPI. In this context, a 27-cm internal diameter multi-coil-array (SCOTCH) was developed [2]. To avoid cross-interactions, a shield must be placed between this device and the RF coil. In this abstract, we propose an 8-channel transceiver array designed to fit inside SCOTCH and made of high-impedance dipoles (HIDs) paired with transmission line baluns, both avoiding the cumbersome use of lumped components.

Based on the same principle as high-impedance loops [3,4], we introduce the HID as a transmission line structure (here as a buried microstrip) with two gaps on its outer conductor making it resonant at the Larmor frequency (499.4 MHz). One can fine tune the HID either by changing the gaps’ positions or their widths. In this simulation, a 6.15 dielectric constant material (Rogers RO4360G) was used for the HID microstrip, making it 26.2 cm long. At 499.4 MHz, the HID is tuned so that the imaginary part of its impedance is equal to 0, and its real part is about 800 Ohms, providing the high-impedance property (Fig. 1A). A transmission line balun (“Marchand balun” [5]) transforms the 800 Ohms to 50 Ohms both required for maximum power transfer at Tx, and for optimal noise matching at Rx (Fig. 1B). A transmit-receive (T/R) switch (Fig. 1C) is integrated into the stripline structure, using three diodes. In the transmit mode, the diodes are biased and protect the preamplifier input (34 dB isolation). In the receive mode, a close to λ/2 long transmission line transforms the low-input impedance of the preamplifier (1.5 Ohms) to a high impedance at the balun’s input, then again transformed into a low-input impedance at the HID port, ensuring preamplifier decoupling between the elements. The stripline structure embedding the balun and the T/R switch is duplicated 16 times with a 20° azimuthal shift. Eight HIDs are connected to every other balun structure; the leftover structures merely close the RF shield and ground plane for the coil. At the front, a copper foil is added to electrically close the shield while saving some space to include a mirror for visual stimulation. Using HFSS (Ansys, PA, USA), the coil was simulated with two homemade multilayers head and shoulders models (Fig. 2). The elements were geometrically fine-tuned on the female model, and no adjustment was done after changing the model to the male one. For completeness, two different HFSS simulations were performed: one at the Tx port, and one at the Rx port (Fig. 1B), both taking into account the parasitic diode effects: a 0.4 pF capacitance in the receive mode, and a 0.4 Ohm resistance in the transmit mode. In each case, the ports were driven with 1W input power. The H-field maps were exported with a 5-mm voxel resolution and transformed to B1+ and B1- profiles. In the receive mode, the S matrix was exported to compute the SNR [6]. Virtual Observation Points (VOPs) were computed to run SAR-constrained kT-points [7] pulse design for different flip angles (maximum average power per channel = 6 W, maximum 10g-SAR limit = 20 W/kg).

In the transmit mode, all elements are matched to lower than -10 dB, ensuring no excessive reflection (Fig. 3A); in the receive mode a low noise correlation is measured in each case and can be attributed to the distance between dipoles (Fig. 3B). The sum-of-magnitudes and SNR maps both exhibit a fairly good homogeneity, except at the top of the brain, which lacks some signal. This signal dropout in this region is more important for the male model and translates into some B1+ shimming difficulties using pulse design (Fig. 4).

A perpendicular-to-B0 shield electrically connected to the baluns’ ground plane at the back of the coil should help to retrieve some signal in the upper brain regions. Even though not yet available at our 11.7T system, a 16-channel transceiver array would certainly allow to further mitigate B1+ inhomogeneities. In that case, a z-segmented array could also be beneficial to give more freedom to the pulse design optimization.

A complete simulation model of an innovative transceiver array at 11.7T was presented. The newly introduced high-impedance dipole paired with a transmission line balun allowed to suppress bulky lumped components. This transceiver array will be later on paired with a tight-fitting cap receive array to maximize SNR at the periphery of the brain [4].

Magnetic resonance spectroscopy (MRS) and tomography (MRT) are valuable techniques for analyzing chemical and biological samples but are often only available to specialized laboratories due to their high cost and complexity [1,2]. To overcome these barriers, it is needed need to develop more cost-effective and accessible NMR technologies [3]. One approach is the use of Software-Defined Radio (SDR), since it offers a promising solution to increase the flexibility and efficiency of NMR analyses [4]. It was also shown that combining pulsed and continuous wave NMR with active RX/TX-switching can improve the signal-to-noise ratio [5]. To develop a more cost-effective solution and to reduce complexity we present a MR signal reception path based on commercial (SDR) sticks, commercial preamplifier, and a custom-built active transceiver (RX/TX) switch tested with a low-field MR system.

For the measurements, an MR signal reception path consisting of a commercial SDR USB stick (NESDR SMArt v5 SDR, Nooelec Inc., Wheateld NY, USA), commercial broadband preamplifier (827becxfh, Walfront, China) and a custom-built active RX/TX switch was developed. The coil, magnet, and transmit pulse generator from a low-field benchtop MR system (MagSpec 0.57 T & Drive L, PureDevices GmbH, Rimpar, Germany) were taken for the experiments. The TX pulse is transmitted through the RX/TX switch into the coil, while the RX signal is routed through the RX/TX switch, passed through the preamplifier, and then directed into the SDR stick. The SDR stick is controlled using GNU-Radio [6], while the TX signal generation and RX/TX switching are managed via benchtop MR in MATLAB software (MathWorks Inc., Natick, MA, USA). The measurement setup is shown in Figure 1. The RX/TX switching is achieved using two PIN diodes. In TX mode, a direct current is applied to PIN diodes D1 and D2, forming a parallel LC circuit with the low-pass pi-circuit components (C1, L, C2). D2 short-circuits C2, preventing strong TX signals from reaching the preamplifier. In RX mode, both diodes are reverse biased, suppressing noise from the power amplifier. Figure 2 shows a schematic of the RX/TX switch. The TX/RX-switch was analyzed with a vector network analyzer (ZNL3, Rohde & Schwarz GmbH, Munich, Germany). To evaluate the MR signal reception capabilities of the SDR stick, a non-selective multi-echo T2 sequence was employed with following parameters: Rectangular Pulses, TE: 4 ms, 38 echoes. During the sequence, raw data containing both (suppressed) TX and RX signals were acquired using GNU-Radio. This raw data was subsequently analyzed in MATLAB, and the T2 relaxation curves were plotted. The experiment was conducted using tap water, a 0.25 mmol/mL contrast agent solution (gadofosveset) in water, and an oil phantom (PureDevices).

Our signal reception setup successfully captures signals, confirming its basic functionality, although signal quality is affected by the lack of electromagnetic shielding, which introduces noise. The impedance of the TX/RX switch matches 50 Ω while in RX mode / PIN diode switched off (S11: < - 60 dB) and low transmission loss (S21: - 0.28 dB) and is mismatched while in TX mode / PIN diode switched on (S11: - 0.58 dB) and high transmission loss (S21: - 31.15 dB). S11 was measured from coil and preamplifier port, S21 from preamplifier to coil port. Both return and transmission loss and the corresponding Smith chart are shown on Figure 3.

The T2 values amount as, 930.16 ± 38.21 ms for tap water, 96.95 ± 1.03 ms for the CA solution, and 144.58 ± 17.12 ms for the oil phantom. These values agree with values measured solely with low-field MR system (tap water: 915.0 ms, CA solution: 97.67 ms, Oil: 141.5 ms) and literature values at 0.5 Tesla, which range for water (500-900 ms), CA solutions (40-100 ms), and oil (50-200 ms) respectively [7,8]. These consistencies validate the setup's accuracy. For improvement, we suggest fully integrating an SDR transmission stick into GNU Radio, allowing cost-effective hardware use while relying on a single, user-friendly software platform. This optimization could reduce development costs and expand potential applications.

This project demonstrates that a functional MR signal reception path can be built affordably by combining cost-effective hardware with open source, user-friendly software which was confirmed by initial proof-of-principle T2 measurements. The required software, GNU Radio, is freely accessible and easy to operate, making this setup accessible even for users with limited technical expertise. This approach opens new possibilities for low-cost NMR applications, making the technology more accessible even beyond the MR community.

Multichannel transmit systems in MRI at 7 T improve excitation homogeneity [1,2], preferring dipoles [3] and meander elements (ME) [4] as transmit elements. Implemented as local coils, they require considerable scanner space. To preserve patients and equipment space, an integrated 32-channel array [5] was installed at the DKFZ. Newer Siemens 7T MRI systems, however, employ smaller gradient coils, reducing installation space by two-thirds. This constrain affects antenna performance, as the H-field magnitude depends on RF shield height [6]. This study evaluates the impact of limited space on transmit capabilities, comparing the performance of ME of the current 32-channel system at different heights over the RF shield with a modified ME and bent dipole (BD) [7].

Modelling and EM simulations were done using CST Studio Suite (CST AG, Darmstadt, Germany). The transmit elements were designed for a Siemens Magnetom 7T system (Siemens Healthineers, Erlangen, Germany) equipped with a SC72 gradient coil at 12 mm distance to the bore liner. The original ME was simulated at 18 mm distance to the RF shield (h₁) and at 9 mm (h₂), to meet spatial constrictions, along a modified ME and BD designs (Fig. 1). Central feeding was co-simulated using a λ/2 balun (330 mm length, 50 Ω impedance) and a matching network with lumped capacitors. All resonant elements and RF shields were simulated as PECs. A lossy 1-mm thick RO4350B substrate was used. The elements were placed below a phantom (330x310x90 mm³) of tissue-simulating liquid (ε'r: 46, σ: 0.87 S/m) at distances of 20 mm (d₁) and 150 mm (d₂). Power (B₁⁺/√Pacc) and SAR efficiency (B₁⁺/√SAR) inside the phantom and coupling were evaluated, with results normalized to 1 W (RMS) of accepted power.

Table 1 shows antenna dimensions and S₂₁ coupling under varied loading. BDs required a small capacitive area for a smaller Lc to maintain central field distribution. At d₁, coupling remained near −20 dB for all elements, while it decreased at d₂ with reduced RF shield spacing. The original ME placed at h₁ showed the highest coupling (−4 dB), reduced to −9.89 dB at h₂. The modified ME had the lowest coupling, reducing S₂₁ by 50% compared to unmodified ME. All BDs maintained <−10 dB values. Contrary to prior work [6], shield proximity improved overall efficiency at short phantom distances (Fig. 2). BD showed higher B₁⁺/√Pacc and B₁⁺/√SAR peak values than the MEs, despite reduced z-axis coverage. At d₂, the element farthest from the RF shield showed superior power efficiency, opposite to the trend at d₁. In general, MEs designs outperformed BD, showing 33% lower B₁⁺/√Pacc. Antennas near the RF shield performed better in superficial phantom layers at d₁ (Fig. 3). Beyond 60 mm depth, all designs performed similarly. At d₂, MEs exhibited superior power efficiency, with no significant SAR efficiency variations. Reflections at the phantom edge caused minor efficiency curve fluctuations due to phantom size.

The impact of RF shield spacing on transmit performance was evaluated for ME and BD at 7 T. In MEs, longer meanders increase electrical length, requiring lower end capacitance (Ce) for compensation [4]. In the ME with short meander, reduced Ce was also required to maintain central field distribution, suggesting a combined effect of increased electrical length from closer RF shield proximity and a thinner microstripline. At larger phantom distances, MEs located farther from the RF shield showed higher power and SAR efficiency. As demonstrated in earlier work [6], increased shield spacing improves field penetration. However, the opposite trend was observed with a close load, suggesting that load proximity also influences field behavior. RF shield proximity improved transmission performance in superficial layers of the phantom, while efficiencies beyond 60 mm depth remained similar. At d₂, ME greatly outperform BDs. SAR efficiency remained similar within the phantom, proving to be independent of shield distance. Coupling was reduced for d₁ due to EM field concentration between coils and phantom, limiting field interactions with neighboring elements. At longer load distances, reduced RF shield spacing significantly reduced coupling across all elements. Lower coupling in all BDs indicated less power transfer to adjacent elements, due to lower B1+ efficiency.

Spatial limitations in MRI systems requires evaluation of transmit elements placed closer to the RF shield for next generation transmit arrays. ME demonstrated strong performance under all simulated conditions. At short phantom distances, MEs near the RF shield demonstrated both improved power and SAR efficiency, while larger shield spacing is preferred when transmitting at greater load distances. Intrinsic decoupling was overall superior for thinner ME. No gain was observed by using distributed capacitors for BDs. Ultimately, ME proved to be a strong candidate for implementation into arrays for future MRI multichannel transmit systems.

At 7 T, multichannel transmit arrays improve B₁⁺ homogeneity [1,2] typically using local transmit coils, which require substantial bore space within the scanner bore liner. To preserve space for patients and equipment, an integrated 32-channel array was installed at the DKFZ [3]. Newer Siemens 7T MRI scanners employ smaller gradient coils, reducing installation space by two-thirds. Reduced spacing between transmit elements and RF shield alters H-field distribution [4], affecting transmission performance. To enable array integration under these constraints, structural modifications to the bore liner constructed from fiber-reinforced plastic (FRP) were explored. Mechanical analysis focused on stress distributions, using Maximum Principal Stress (MPS) and von Mises Stress (SvM) [5], criteria suited to evaluate structural response in FRPs [6]. Alternative bore liner designs with integrated antenna slots were analyzed using finite element analysis (FEA) under conservative loading.

The modelling and finite element analysis (FEA) structural simulations were done with ANSYS Mechanical (ANSYS, Inc., Canonsburg, PA, USA) using a bore liner model from Siemens Magnetom 7 T system (Siemens Healthineers, Erlangen, Germany). Material data was provided by Maschinenfabrik Reinhausen GmbH (Reinhausen, Regensburg, Germany). The bore liner consists of E-glass/epoxy (roving, 79.5 wt.-% glass), with mechanical limits set by IEC 61462: 60 MPa axial and 120 MPa tangential stress. E-Glass Wet was simulated for its similar anisotropic properties. To accommodate the 32-channel array (Fig. 1), 32 antenna-sized slots (250×50 mm²) were introduced with spacing of 28 mm (xy-plane) and 50 mm (z-axis). Load simulations accounted for a combined mass (patient, table, and dedicated receive array) of 2000 N, with 1000 N applied per rail for symmetric loading. Four fixed supports (110×96 mm²) constrained the structure. An adaptive mesh was refined near critical regions. Three designs were analyzed: original, slotted with sharp edges, and slotted with 24 mm-radius edge rounding. For all models, MPS and SvM assessed the static structural response.

Fig. 2 and 3 show simulation stress distribution in the bore liner, viewed in lateral and base sections. Table 1 shows the maximum FEA values obtained for each case. For stress analysis, a MPS of 2.89 MPa was obtained for the original bore liner (A). Introducing 32 slots with sharp-edged corners increased the stress nearly twice compared to the original design, increasing to 7.18 MPa. Applying a 24 mm radius to the 32 slots (C) reduced the value to 5.84 MPa. In SvM calculations, the original structure (A) exhibited a stress value of 2.83 MPa. Introducing 32 slots with sharp corners increased the stress to 6.8 MPa, nearly 2.5 times higher than the original. Applying a 24 mm radius at the slot base reduced the stress to 5.81 MPa, representing a decrease of approximately 1 MPa compared to the sharp cornered design. Across all models, the stress values obtained in both analyses were well below the limits under internal pressure of 60 MPa and 120 MPa. The highest stress concentrations were consistently located near the borders of the supporting rails, with pronounced intensity in the central base area of the structure (Fig. 3 A-F).

The original bore liner consistently exhibited the lowest values across all failure metrics, serving as a comparison baseline. Introducing sharp-edges slots significantly increased stress, nearly tripling peak failure values due to geometric discontinuities acting as stress concentrators. These results align with previous work [5,7], showing that in FRPs stress concentrates around cut-out areas. Rounding the slots edges notably reduced peak stress, indicating improved load distribution, resistance to failure and reducing the stress concentration factor. Replacing sharp with rounded edges redistributes stress more evenly, consistent with established strength optimization strategies [6]. All stress levels remained below IEC 61462 limits, with a maximum of 7.18 MPa (~10% of max. stress limit), well within the E-glass allowable range. This suggests that the design modifications are likely to maintain structural integrity under the applied load conditions. Top region of the liner showed minimal stress (in contrast to stress concentration along rail borders), suggesting it as a viable area for RF coil integration under reduced mechanical loading.

High-count integrated arrays are incompatible with the spatial constraints of new Siemens 7 T systems. Structural modifications to the bore liner were evaluated to enable array integration and preserve performance while maintaining mechanical stability. FEA simulations showed that refining the slot geometry with rounded transitions effectively decreased principal stresses concentrations and improved load distribution. These findings support the feasibility of slot integration, provided that key geometric features are optimized for mechanical stability.

Impedance matching plays a critical role in optimizing the performance of radiofrequency (RF) coils in Magnetic Resonance Imaging (MRI), especially under low magnetic field conditions where the signal-to-noise ratio (SNR) is inherently limited[1, 2]. Traditional matching circuits rely on variable capacitors, which are not only sensitive to loading conditions but also contribute to noise and increase setup complexity. In this project, we explore an alternative, passive method of impedance adaptation using loop inductive coupling to simplify the matching procedure of a solenoid coil. The objectives are component cost and complexity reduction, and SNR improvement, in the context of 0.1 T MRI exploration.

Experiments were conducted on an open 0.1 T vertical-field electromagnet (EAR 50 L, Drusch, France) (Fig. 1). This resistive magnet is powered by a high-stability power supply and water-cooled during operation. This system also embeds magnetic field gradient coils and shimming coils (B0 homogeneity correction on 3 first order and 5 second order components). The transmit-receive solenoid coil[3] consists of a 12 turns solenoid (125 mm x 65mm) made with 6 mm cross-section copper tube. It is tuned to the Larmor frequency thanks to three parallel 22 pF ceramic capacitors and two 0.8 – 8 pF variable capacitors directly soldered onto it. To facilitate precise tuning prior to acquisition, resonance is achieved by adjusting the B0 field strength through modulation of the magnet power supply current. A secondary non-resonant inductive loop, mechanically repositionable along the solenoid long axis, is used to adjust impedance matching instead of an additional circuit as traditionally done[4]. Matching efficiency was evaluated using a Vector Network Analyzer (ZNB4, Rohde & Schwarz). Two loop positions were compared: Pos0 = the loop is positioned at the solenoid entrance, and Pos1 = the loop is positioned around the solenoid at a location minimizing antenna’s return loss - S11 (dB). S11 and quality factors were measured with and without the phantom load (a cylinder filled with tap water mixed with copper sulfate (CuSO₄) to shorten relaxation times). MRI data were acquired on the phantom using three different sequences: 2D Gradient Echo (TE = 3.8 ms, TR = 202 ms, FA = 60°, FOV = 250mm2, matrix = 64 x 64), 2D Fast Spin Echo (TE = 5.06 ms, TR = 500 ms, ETL = 5, FOV = 250mm2, matrix = 64 x 64), and 3D Spin Echo (TE = 5.78 ms, TR = 500 ms, FOV = 250mm2, matrix = 64 x 64). NMR signal profiles, noise background and nutation curves were obtained with standard one pulse sequence and were processed with MATLAB. The mean signal and the noise standard deviation were extracted from regions of interest defined using the graphical tool integrated into the system console (Chameleon 4, RS2D, France). The SNR was then calculated by accounting for the Rayleigh distribution of noise in magnitude images[5] as follows: SNR= Mean(Signal)/(SD(Noise))× √(2- π/2) (1)

At Pos1, the loaded coil achieved a minimum S11 of –40.68 dB at 4.54 MHz vs -8dB at Pos0 (Fig. 2). Q factor calculated as the ratio between Qunloaded and Qloaded was 1 in both cases. NMR signal peak showed good field homogeneity in both cases (<13 ppm). Signal intensity of one pulse acquisition increased by 59.4% in Pos1 compared to Pos0 (Fig. 3.A). Standard deviation of noise acquisition also increased by 27.6% though. Nutation experiments confirmed same excitation pulse calibration led to better signal reception in Pos1 (Fig. 3.B). Image acquisitions showed significant enhancement in signal intensity at Pos1, with decreases or only minor increases in background noise depending on the evaluated slice, resulting in an overall improvement of SNR across all slices for the three imaging sequences (Table 1).

The use of an inductively coupled loop for impedance matching presents multiple advantages for low-field MRI systems. It eliminates the need for noise-prone electronic components, enables rapid and user-friendly matching adjustments, and reduces development costs. Additionally, this method allows flexible adaptation of coil configurations to varying subject morphologies or phantom geometries that lead to different filling factors, improving overall signal capture. The measured SNR gains demonstrate the method’s clinical potential aligning with current trends in developing lighter and cost-effective low-field MRI systems.

We demonstrate the feasibility of achieving highly efficient and flexible impedance matching for a solenoid coil using passive inductive coupling in a 0.1 T MRI system. This approach significantly enhances SNR without electronic components, reduces system complexity, and opens new perspectives for the development of accessible point-of-care MRI technologies.

Combining proton(¹H), sodium(²³Na), and potassium(³⁹K) imaging offers significant clinical potential for investigating Na/K pump deficiencies associated with conditions like epilepsy, stroke, or cancer(1–3). While triple-frequency coils(¹H/²³Na/³⁹K) have been demonstrated for preclinical application(4), no equivalent for human has been reported to date. Previously, we presented separate coil designs: a 16-channel ²³Na-loop/¹H-dipole array(5), and a ³⁹K birdcage coil(6) for 7T MRI. In this study, we introduced a novel triple-resonant RF coil for human head imaging, integrating our previously developed coils. EM simulations were performed to optimise coil structure in terms of S-matrices and B₁⁺ performances and assess RF safety with SAR₁₀g using a human head model.

The triple-frequency coil combines previously developed 16ch ²³Na-loop/¹H-dipole array(Fig.1a) and ³⁹K birdcage coil(Fig.1b). The ³⁹K birdcage was inserted into the existing ²³Na/¹H structure, forming a double-layered coil. Four orientations were simulated to identify the optimal configuration(Fig.1c): 1.0° rotation-birdcage sources aligned symmetrically with dipoles;2.22.5° rotation-birdcage sources placed midway between dipoles;3.45° rotation-birdcage sources realigned with dipoles;4.dipoles directly mounted onto the birdcage instead of the loop layer. Finite-difference time-domain(FDTD) simulations(Sim4Life 8.2,ZMT,Switzerland) were performed with a spherical phantom(d=15 cm,εᵣ=81,σ=0.95 S/m). Dipoles were modelled as centre-shortened structures on FR-4 substrates(εᵣ=4) with lossy metal(σ=5.8×10⁷S/m). Loop coils and birdcage elements were modelled as perfect electric conductors (PEC) with zero thickness surfaces. Alternate loop coils were elevated by 8 mm to facilitate voxelisation. An RF screen(PEC) was also included. Individual nuclei simulations (¹H:297.2 MHz, ²³Na:78.6 MHz, ³⁹K:13.9 MHz) were performed separately, with other coil elements present but inactive. Co-simulation (OptenniLab 5.2,Finland) adjusted lumped element values for tuning and matching. S-matrices and circularly-polarised(CP) B₁⁺ maps were extracted to assess coil performance (Figs.2,3). Separate simulations were performed for single-nuclei elements for comparison. Following phantom optimisation, RF safety was evaluated using the Duke head model at 2-mm iso resolution(Fig.4a). B₁⁺ and SAR₁₀g maps were interpolated to 1-mm iso resolution, normalised to 1W total input power, and exported to Matlab for analysis.

S-parameters for all setups were compared to independent single-nuclei coils (Fig.2). Setup 4 (dipoles on birdcage) provided optimal 1H performance, achieving the lowest reflection (Sᵢᵢ=-25.6 dB) and best decoupling (Sᵢⱼ<-13.4 dB). ²³Na and ³⁹K coil performances were consistent across setups, closely matching individual coil benchmarks. B₁⁺ maps indicated significant orientation dependence for ¹H imaging(Fig.3). Setups 1–3 demonstrated reduced B₁⁺ efficiency(17–54% decrease in mean B₁⁺) compared to independent ¹H-dipole array, while Setup 4 markedly improved mean B₁⁺ by 66.7%. For ²³Na, all setups showed increased B₁⁺ of 43–68% with minimal orientation dependence. The ³⁹K coil showed consistent increase of 17–20% in B₁⁺ across all setups. Due to superior performance across all nuclei, Setup 4 was selected for safety evaluation. Duke head simulations confirmed good matching(Sᵢᵢ<-15 dB) and isolation(Sᵢⱼ<-7 dB). B₁⁺ on the central slice and maximum intensity projection(MIP) for SAR₁₀g are shown in Fig 4. Maximum SAR₁₀g values were reported as 0.25W/kg(¹H), 0.12W/kg(²³Na), and 0.24W/kg(³⁹K).

In this study, we proposed a compact triple-resonant coil integrating ¹H, ²³Na, and ³⁹K imaging elements at 7T. Setup 4 provided the optimal balance of coil performance, demonstrating excellent S-parameters and efficient B₁⁺ fields. This advantage was most pronounced for ¹H, likely due to dipoles mounted directly on the birdcage, reducing shielding effects in other setups. Notably, setups 1 and 3 yielded similar B₁⁺ magnitudes but differed in spatial orientation, with the ¹H field pattern clearly rotating by 45° between these configurations. In Duke simulations, the coil showed maximum SAR₁₀g approximately 30% higher for ¹H compared to previously reported values for separate ¹H-dipole array, concentrated mainly near the ears due to their proximity to dipole elements. SAR₁₀g levels for ²³Na and ³⁹K remained comparable to previously validated coil designs, indicating minimal additional risk introduced by combining elements. However, additional safety validation remain necessary with human evaluations.

We have demonstrated a promising initial design and simulation-based optimisation of triple-resonant(¹H/²³Na/³⁹K) RF coil at 7T. The optimised coil(Setup 4) achieved excellent multinuclear performance with SAR evaluated in simulations using a realistic human head model, highlighting its potential clinical utility. Future work will focus on coil fabrication, bench testing, and safety validation.

High-field MRI systems benefit from parallel receive coil arrays, which can significantly enhance the signal-to-noise ratio (SNR). Current state of the art research increase the number of coils as a means to increase the SNR [1-4], where the coil geometry is not optimized. One of the primary challenges in multichannel receive arrays is interelement decoupling. This decoupling is a key limiting factor in optimizing both the individual coil geometry and the overall array configuration for maximum signal-to-noise ratio (SNR). Traditional decoupling techniques, such as partial coil overlap and preamplifier decoupling, inherently constrain the flexibility of coil design. Recently, novel coil designs based on transmission line modifications [5, 6] have been introduced as self-decoupled elements. These coils maintain effective decoupling and performance even when their shape deviates from the conventional circular form, which makes them a good candidate for optimized coil geometry. This study proposes a novel approach to designing receiving coil array geometries using an algorithm to find the optimal coil shape combination that maximizes SNR in user defined regions using a limited number of coils.

420 coils were modeled using 100 line-segments per coil. The coils were placed on a cylindrical plane 2 cm from the phantom. All coils have a wire length of 31.415 cm and consist of various shapes (squares, triangles, vertical and horizontal rectangles, circles and figure eights). A homogeneous cylindrical phantom was modeled (height = 20 cm, radius = 10 cm, σ = 0.55 S/m, ε = 80, ⍴ = 997 kg/m^3) with a voxel resolution of 6mm. The electric field (E) and magnetic flux density (B) were simulated per coil individually in the phantom using the MARIE solver [7] in MATLAB. The loss matrix is calculated as Q = dV ∑ σ(r)E(r)E(r)^H, where dV is the voxel volume. The B1 fields are calculated as B1- = (Bx - 1i*By)/√2 and B1+ = (Bx + 1i*By)/√2. The SNR per voxel is then calculated as SNR(r) = B1- x/√(x^H*Q*x), where x is a weight vector between the coils which is to be found for maximum SNR. The solution to this optimization problem is well known and will not be discussed here. 2 regions are defined in the phantom, the periphery (|z| < 5 cm, r > 8 cm) and the center (|z| < 5 cm, r < 6 cm), as shown in Fig. 2 (a). The optimization goes as follows. First a region of interest is chosen (periphery or center), then the average SNR of all coils is in this region is calculated individually. The one with the highest mean SNR is added to the array. Next all non selected coils are individually tried in combination with the existing array. The one resulting in the highest mean SNR is added to the array. This process is repeated until the array consists of 16 coils (Fig. 1). As a comparison a conventional equispaced 16 coil circular loop array is also examined. Loops were placed in two rows (Fig. 2(b)) and each loop was of 10cm diameter.

The found geometries of the optimized arrays are shown in Fig. 2 (c-d). Calculations of the SNR, portrayed in Fig. 3, demonstrate a significant improvement for the optimized coil arrays compared to circular loop array. The periphery optimized array shows a mean increase of 67% of the mean periphery SNR and a 15% decrease in the mean center SNR compared to the loop array. The center optimized array shows a mean increase of 17% of the mean periphery SNR and a 6% increase of the mean center SNR compared to the loop array.

With the approach shown in this manuscript, we demonstrated that improvement in peripheral SNR can be achieved by optimizing the coil geometry and its position within array. While algorithmic optimization outperforms conventional loops, it does not guarantee a globally optimal solution, an improved algorithm might yield better results. Future research should include algorithm optimization strategies. The assumption of complete decoupling does not hold for all coil types, in this research transmission line coils are assumed. If other coil types are used additional decoupling hardware might be necessary for the SNR to be improved using these coil geometries. Further improvement in SNR can be achieved by considering specific clinical applications, such as optimizing for a specific part of the brain. Future research will focus on realistic human body models. Also, as a future work optimization would be performed on higher channel number such as 32, 64 or even higher.

Algorithmic optimization has been shown to improve SNR at 7T using a 16-channel receive coil array. This approach provides a promising avenue for enhancing image quality in high-field MRI systems.

MRI is a modality of choice for spine imaging as it allows optimal visualization of soft tissues. In the last decades, Ultra-High-Field scanners operating at 7 T have been developed with the aim of improving signal-to-noise ratio (SNR) and resolution. However, these technical advantages came at the cost of new challenges such as increase of the specific absorption rate (SAR) and inhomogeneities of the radiofrequency field B1 [1]. These inhomogeneities are mainly related to the shorter operating wavelength and have changed the concept of optimized RF coils. Indeed, at UHF, volumetric coils might not be optimal and surface coils adapted to a given region of interest (ROI) might be preferred. Our target area, the thoracic or lumbar spine, is 25-30 cm long, less than 1 cm in diameter, and lies at 4-6 cm under the skin [2]. The optimal coil design for this region for UHF MRI is not yet known [3]. We propose a coil array specifically adapted for the thoracic and lumbar spine geometry. It is made of 8 loops that can be independently controlled in amplitude and phase, thus allowing for B1 shimming and parallel transmit [4][5]. While some studies have shown the interest of using active dipoles to improve in depth transmission [6], we provide evidence that B1 intensity and homogeneity in the ROI can be improved by adding carefully placed passive dipoles to our design. In addition, maximum SAR can be reduced thereby ensuring a better transmit efficiency.

Eight 10 cm diameter loops, segmented with three 2.7 pF capacitors were positioned in two rows and overlapped to minimize inter-element coupling. Each loop was tuned and matched to the Larmor frequency of 297.2 MHz and connected to the MRI scanner through an interface box (Stark Contrast, Germany). In addition, two metallic stripes of 600 mm length, 12.5 mm width and segmented in their middle with 10 pF capacitors were placed under the outer edges of the coil array in order to act as passive dipoles. Simulations of this setup were run on the time-domain solver of CST Studio Suite with a Finite Integration Technique. In both numerical simulations and experiments, the array was placed on a foam spacer 10 mm above the 2 dipoles which were located 10 mm above a 480x350x175 mm3 phantom filled with a dedicated body liquid (εr=58,2 and σ=0,92 S.m-1). (Fig. 1) Experiments were conducted at 7T (Magnetom TERRA, Siemens Healthineers, Erlangen, Germany). B1+ maps were acquired with a turboFLASH sequence [7]. A reference static excitation was selected, with each channel set to the same amplitude and a 180° phase shift between the two rows in order to produce constructive currents in the center of the array [8]. The performance of the array and the two dipoles combined was compared to the performance of the array alone. To do so, a region of interest corresponding to the hypothetic position of the spinal cord was defined as a 300x3x3 mm3 box lying 5 cm deep into the phantom.

Simulations indicated that the addition of the 2 dipoles improved the average amplitude of the field by 21% and reduced the relative standard deviation by 34%. The maximum SAR (in the whole phantom) was reduced by 26% and thus the transmit efficiency (defined as the ratio between average amplitude and square root of maximum SAR) was improved by 40% in the ROI. (Table 1) Experiments confirmed that the 2 dipoles were effectively focusing the field in the central region (Fig. 2). The average B1+ field was improved by 25% while the Relative Standard Deviation was reduced by 15% in the ROI (Table 1).

In the case of the array alone, the current patterns applied in the loops to generate a bright B1+ field in the ROI also lead to undesirable illuminations under the outer edges of the array. By adding the two dipole-like elements and tuning them so that their surface currents become out of phase with those of the outer edges of the array, we can offset the lateral contributions and thus cancel the B1+ field outside of the desired area (Fig. 3) The experimental results will need to be reproduced and tested for different ROIs to determine error margins and identify possible causes of mismatch with the simulations.

Our results illustrate that through a better distribution of the transmit field, the addition of passive dipoles can improve the classic coil array design. Simulations indicate a gain in both B1+ homogeneity and intensity, and also predict a reduction of the maximum SAR. Experiments further confirm this working principle and show that the dipoles effectively focus the field on the ROI. The in vivo assessment of this prototype is under progress.

Development of radiofrequency (RF) coils for clinical MRI focuses on enhancing image quality, patient safety, and comfort. Wireless RF coils operating by inductive coupling with the body birdcage coil (BC) are an alternative to conventional cable-connected RF coils [1, 2, 3]. These coils offer significant advantages over traditional cable-connected coils, including reduced electromagnetic interference and improved patient comfort since no thick cables are used. However, since wireless coils are usually operating in transceiver (TxRx) mode, it is necessary to perform manual calibration of the reference Tx voltage to avoid inhomogeneity of the B1+ field [2]. It complicates clinical workflow and increase scan time. To address these limitations, we have developed a concept of a wireless receive-only RF coil (Rx-only) for 1.5T MRI [1], specifically designed for breast imaging where dedicated multi-channel coils are often unavailable across different vendor platforms. The design is based on a Helmholtz coil with two passive diode traps for detuning. This means that imaging setup with wireless coil is performed in Tx mode similarly to traditional setup with cable connected Rx coils. In receive mode, the Rx-only wireless coil is tuned and is able to detect the MR signal, which then is injected to Rx-chain of MR scanner via inductive coupling with BC. Therefore, it makes possible to use this Rx-only coil without specific adjustments [1, 2]. In this study, we evaluate the multi-vendor compatibility of the proposed Rx-only wireless coil through experimental validation with 1.5T (Larmor frequency 63.8 MHz) MRI systems of major vendors (GE, Siemens, Philips). Results demonstrated the coils ability to perform MR imaging without disrupting standard imaging protocols, offering a practical solution for clinical MRI across different vendors.

The experimental study was performed using our previously developed Rx-only coil prototype [1]. This is a Helmholtz coil incorporating two passive LC traps (with inductance of 67 nH and a capacitance of 89.5 pF) inserted into the gap of the coil conductor (Fig.1d). Each inductance is connected in series with a pair of back-to-back Shottky diodes. During the transmit mode, forward-biased diodes activate the traps, shifting the resonance frequency away from the Larmor frequency. In the receive mode, reverse-biased diodes disable the traps, that allows to operate Rx wireless coil in receive mode. Experimental study was performed on three different 1.5T MRI systems using the same spherical phantoms filled with a NiCl2 water solution. Siemens Espree Magnetom was used to acquire phantom images and flip angle (FA) distributions using the double-angle method [4] (TR/TE = 9000/4.76 ms, matrix = 128×128 mm²). For Siemens system SNR calculations were performed with FA = 90°. FA mapping employed gradient-echo sequences with FA of 40° and 80° for the double-angle reconstruction. GE Optima 360 system was used for comparative analysis of the Rx-only coil versus 8-channel HD Breast coil using identical scan parameters. On the Philips Achieva system, qualitative image assessment was performed against a SENSE breast 7-channel coil. All experimental setups for each vendor platform are shown in Fig. 1 a-c.

Figure 2a shows the phantom image acquired with the Rx-only coil on the Siemens system, demonstrating an SNR of 650. In comparison, the BC-only configuration achieved a lower SNR of 115. For the GE scanner, the phantom image acquired with the Rx-only coil (Fig. 2b) was compared with 8-channel HD Breast coil (Fig. 2c). Additionally, Philips system evaluations of the wireless Rx–only coil (d) and 7-channel SENSE coil (e) are presented. On GE and Philips systems it is not possible to disable the transmit coil by software, so it is not possible to obtain correct noise maps for SNR assessment. Figure 3 presents the FA distribution maps obtained from Siemens scanner for: (a) the Rx-only coil and (b) BC-only configuration.

The results demonstrate successful multi-vendor compatibility of the Rx-only coil across three MRI vendors. The design overcomes vendor-specific limitations like absence of manual voltage calibration required by Tx/Rx wireless coils, particularly critical for GE and Philips systems. The wireless coil provides consistent performance without protocol modifications and has much lower cost.

The wireless receive only coil provides multi-vendor compatibility with 1.5T MRI-systems eliminating manual calibration of Tx voltage. This design offers a cheaper and simpler alternative to conventional Rx coils. - This work was supported by state assignment No. FSER-2025-0018 within the framework of the national project “Science and Universities” -

Conventional surface coils in MRI are inherently limited to a field-of-view (FOV) constrained by their physical dimensions. The traveling-wave MRI (twMRI) approach overcomes this limitation by enabling large-FOV imaging, as previously demonstrated [1]. In this work, we present an enhanced twMRI system combining: (i) a parallel-plate waveguide (PPWG) that eliminates cutoff frequency restrictions, and (ii) a bioinspired surface coil design [2]. Building on prior work showing metamaterials can improve signal-to-noise ratio (SNR) in this configuration [3], we address a fundamental limitation of traditional twMRI implementations. The primary objective of this study was to experimentally validate that our twMRI-PPWG system with a bio-inspired surface coil can achieve an extended FOV beyond the coil's physical dimensions while maintaining high image quality.

A bio-inspired surface coil, as described in [2], was employed for RF signal transmission. The metamaterial structure, designed in a nonagonal configuration, consisted of nine copper strips (500 mm × 10 mm × 35 μm, σ = 5.96 × 10⁷ S/m) laminated onto FR4 substrates (ϵ = 4.35, tanδ = 0.008, 1 mm thickness). Each strip featured a periodic array of 49 C-shaped unit cells (5 mm diameter, 3 mm gap width, 2 mm conductor width; Fig. 1.b) arranged linearly in a 1×49 configuration (Fig. 1.a). Phantom imaging experiments were conducted using a saline-filled parallel-plate waveguide (PPWG) with the metamaterial positioned inside (Fig. 1.c). The bio-inspired surface coil was placed externally, aligned parallel to the waveguide plates to maximize coupling. To validate the approach, sagittal phantom images were acquired with a standard gradient-echo sequence (TE/TR = 4.39/200 ms, FOV = 60 × 60 mm², matrix = 256 × 256, flip angle = 45°, slice thickness = 1 mm, NEX = 4). For comparison, additional images were obtained using both the bio-inspired coil and an in-house birdcage coil (length/diameter = 5 cm/4 cm, 4 rungs). All experiments were performed on a 7T/30 cm Bruker scanner (Bruker BioSpin MRI GmbH, Germany) and cylindrical phantom (diameter/length = 4 cm/ 10 cm) filled with saline solution.

Phantom images acquired with the nonagonal metamaterial confirmed the feasibility of the approach (Fig. 2b). Signal-to-noise ratio (SNR) analysis yielded the following values: SNR(nano) = 99, SNR(inhBC) = 106, and SNR(bio) = 115. A profile comparison of the twMRI and the in-house BC coil data (Fig. 2.a) revealed similar performance between the two methods. However, the twMRI signal exhibited a gradual decline with distance from the coil, reaching a maximum decrease of 27% relative to the initial signal.

Notably, the twMRI image avoided the pronounced signal drop observed at the end rings of the in-house BC coil profile (Fig. 2.a), demonstrating a key advantage. Furthermore, the phantom images (Fig. 2.b–c) indicated that the twMRI provided greater coverage than the bio-inspired surface coil. This is a rather encouraging result, as the signal was transmitted 58 cm away from the phantom's location.

Our results demonstrate that the twMRI approach achieves a significantly larger FOV than conventional surface coils, exceeding the coil’s physical dimensions while maintaining competitive image quality. This work represents a key advancement in the practical implementation of twMRI, enabling expanded anatomical coverage without compromising sensitivity—a critical requirement for many imaging applications.

Over the past decade, there has been an increasing interest for radiofrequency devices dedicated to imaging of specific anatomical areas, such as breast, wrist etc... Metamaterial resonators at 1.5T [1,2] and ceramic resonators at 3T [3] have already been shown to be effective for targeted MRI applications but experimental data related to a metamaterial volumetric resonator operating at 3T are missing. Here we propose a passive metamaterial resonator designed for a 3T MRI scanner. In addition, we analyzed the influence of two birdcage excitation modes.

The design of the metasolenoid illustrated in Figure 1 was inspired by a previously reported volumetric resonator operating at 1.5T [4]. It consists in four PCB plates soldered together in a rectangular box shape. Each plate contains 10 etched copper strips so that the entire device is composed of 10 split-loop resonators, with a gap on each plate for capacitors. Its resonance frequency was tuned to 123.2 MHz by adding three 3.3 pF and one 2.7 pF ceramic capacitors (Johanson Technology and Knowles) in the gaps of each SLR. The resonance frequency was also characterized by connecting a non-resonant loop antenna connected to a vector network analyzer (MS2026C, Anritsu). A tissue-mimicking liquid (MVG, France) with a relative permittivity of 61.9 and a conductivity of 0.8 S/m was used to create the phantom by filling a 2.4 L glass bottle. The experiments were performed in a 3T Vida MRI scanner (Siemens Healthineers, Erlangen, Germany) using a whole-body birdcage coil (BC) for both signal transmission and reception. Flip angle (FA) maps were acquired in the sagittal orientation using a pre-pulse RF with a TFL readout sequence [5], TR/TE = 7030/1.8 ms, FA = 8°, Sinc pulse = 80°, FOV = 289 x 289 mm, matrix size = 128 x 128 and 13 slices of 10 mm thickness. The reference voltage was calibrated for each set of measurements. The birdcage excitation was used as a reference. We performed measurements with two different polarizations : circular polarization (CP) with both birdcage ports excited in quadrature and linear polarization (LP) by selecting only the birdcage port that couples to the metasolenoid resonator.

The reflection coefficient shown in Figure 2 was measured using a non-resonant measurement loop in order to estimate the resonant frequency of the metasolenoid. A sharp dip could be observed near the Larmor frequency when the metasolenoid was not loaded, indicating that the metasolenoid was properly tuned. Accordingly, for the loaded metasolenoid, a reduction of the resonance quality factor was observed. The experimental setup and FA maps are shown in Figure 3. Panels B–E show the FA map in the mid-slice of the phantom. The FA averaged in the ROI and the reference voltage are given for each measurement. Panels B–C show the Birdcage reference coil for CP and LP excitations, respectively, panels D–E show the results with the meta-solenoid. Table 1 presents a summary of all results. As compared to the reference birdcage-only experiment, the addition of the metasolenoid was able to significantly reduce the reference voltage required to achieve the target FA in the phantom ROI. Transmission efficiency was calculated as the ratio between the averaged FA in the ROI and the reference voltage. Using the circular polarization, a 3.2-fold transmission efficiency increase was measured. A much larger i.e. 5.8-fold increase was measured using the linear BC polarization.

The present results indicate that a metasolenoid passive resonator can largely improve transmit efficiency in a 3T MRI system. The corresponding enhancement factors were comparable to those expected by simulations [6]. Of interest, transmission efficiency was improved differently according to the birdcage polarisation. More specifically, using the BC linear excitation led to a further reduction of the input power required to reach a given FA thereby illustrating an improved transmission efficiency. When comparing results from each polarisation, we observed a +26% improvement of the transmit efficiency with the LP which has not been previously reported in the literature. In other words, the optimized response from the resonator was obtained when the BC port was excited so as to generate a magnetic field collinear to the metasolenoid axis. As illustrated by the results obtained with the CP, the metasolenoid did not interact with the BC when the magnetic field was generated in a perpendicular plane.

The metasolenoid resonator designed for the present study was able to largely improve the magnetic field transmit efficiency within a targeted region. A three-fold increase was obtained with a quadrature driven birdcage. Of interest, the transmit efficiency was further improved (six-fold) for the linear driven birdcage. Additional numerical SAR studies will be conducted in order to confirm the benefits of the metasolenoid approach in vivo.

Ultra-high-field (UHF) MRI holds great potential for brain studies, offering SNR and CNR compared to clinical (1.5 and 3T) MRI [1]. At 7T (297 MHz), the shorter wavelength (~10 cm in the human body) results in severe field inhomogeneities and standing-wave effects, limiting uniform radio frequency (RF) coverage. More than a decade ago, the traveling-wave (TW) MRI method was introduced [2]. This method avoids local transmit (Tx) coils which usually requires multiple channels and complex magnitude/phase configurations. However, TW MRI has a significant drawback of low B1+ (Tx-efficiency) [3]. Several methods have been proposed recently to improve the efficiency and homogeneity of TW MRI excitation using additional passive structures: a coaxial waveguide for brain imaging [4], high-permittivity dielectric as a bore liner [5] or a local dielectric waveguide (DW) surrounding a human head [6]. However, all these structures are bulky and causing aliasing due to water signal. In this work, we propose a thin and light cylindrical metasurface (MS) represented with a flexible PCB to improve the efficiency of TW MRI at 7T that exhibits the Tx-efficiency similar to one of a DW [6].

To achieve an equivalence between a MS and a dielectric slab (DS) (th=28 mm, ε=52), the MS unit cells were designed to mimic the DS's slow-wave factor (for the Tx field localization effect). To reach equivalence, the phase shift through the dummy parallel-plate waveguide section containing one chain of the MS unit cells (a cross of сopper strips loaded with lumped capacitors C) must match that of the same length waveguide section with the DS fraction. Numerical models of MS and DS waveguides are shown in Figure 1A,B. The optimal MS unit cell (period of 17 mm, strip width of 0.3 mm) required 3.9 pF capacitance. A full model of the cylindrical MS surrounding a human head model (Figure 2A) was built upon a previously optimized single-element design. In the practical realization, copper strips of the MS are assumed be made on two sides of a thin and flexible polyimide PCB (th=0.05 mm, ε=3.5) ending with parallel-plate printed capacitors (square side of 2.4 mm) (Figure 2B) as used in [7]. The size of the parallel-plate capacitors to replace the lumped ones was determined using a phase delay comparison in a waveguide at 297 MHz. The B1+ field was calculated for the MS loaded with the Duke voxel model [8]. A CP patch antenna similar to [6] was placed at a 140 cm distance from the top of the voxel model to provide TW excitation. The complete numerical model is presented in Figure 2C. The cylindrical MS design was optimized by varying its length, radius of the cylinder and axial displacement relative to the head model to achieve optimal B1+ homogeneity and Tx-efficiency. B1+ field homogeneity was assessed as coefficient of variation (COV) of B1+ across a 180-mm-thick transverse slab that includes the entire brain. The SAR was also calculated using the CST Legacy averaging method over 10 g of tissues. The SAR-efficiency was defined as the ratio of the mean B1+ to the SAR level. For comparison, the DW from [6] also were simulated in the same TW setup.

The correspondence between the MS and the DS was evaluated using the relative slow factor (RSF) shown in Figure 1C. The best coincidence occurred at an RSF value of 0. The frequency-dependent RSF reached a minimum of 297 MHz, demonstrating that the optimized MS provided the same phase delay as the DS at the Larmor frequency of a 7T scanner. Figure 3A presents B1+ in the central sagittal plane of the model obtained using different configurations. Figure 3B shows the parametric results obtained in the presence of the proposed MS, including the B1+ homogeneity, Tx-efficiency, pSAR10g and SAR-efficiency values calculated for a fixed MS diameter of 280 mm and various lengths. Figure 4A shows numerically simulated B1+ in the central sagittal slice using the optimal configuration (410 mm length, 280 mm diameter) of the MS and the DW.

As seen in the table (in Figure 3B), the MS with a length of 410 mm provides the best compromise between the COV, Tx-efficiency, pSAR10g and SAR-efficiency. As shown in Figure 4B, the proposed MS substantially improves the RF field distribution compared to the DW. Importantly, the proposed MS provides substantially better B1+ homogeneity (by 0.4%), Tx-efficiency (by 22.3%), pSAR10g (by 7.5%) and SAR-efficiency (by 27.1%) compared to the state-of-the-art DW, while also possessing advantageous features such as reduced thickness and weight improving the patient’s comfort.

In this work, we numerically investigated a novel cylindrical MS for human head TW MRI at 7T with improved efficiency and homogeneity. The MS geometry is based on a grid consisting of copper strips loaded with parallel-plate capacitors printed on a flexible polyimide substrate. The proposed MS can be used in 7T MRI applications, where convenient access and high Tx-efficiency are required simultaneously (e.g., in fMRI).

The traveling-wave MRI (twMRI) approach using a parallel-plate waveguide (PPWG) enables imaging with a larger field of view without cutoff frequency limitations [1]. Previous studies have shown that this scheme yields images with reasonable uniformity but lower signal-to-noise ratio (SNR) values [2]. However, incorporating a metamaterial-loaded PPWG has been demonstrated to improve SNR [3]. In this work, we propose an alternative approach: a double-layer metamaterial-loaded PPWG filled with saline solution and integrated with a transceiver bio-inspired surface for preclinical imaging.

A bio-inspired surface coil was employed for RF signal transmission, as described in [1] (Fig. 1c). The metamaterial consisted of two arrays of 3 × 50 C-shaped units fabricated from copper sheets (thickness = 35 microm, s = 5.96 × 10⁷ S/m) laminated onto an FR4 dielectric substrate (e = 4.35, tan(d) = 0.008, thickness = 1 mm, dimensions = 500 mm × 40 mm), forming a sandwich-like structure (Fig. 1b). Each C-shaped unit had a diameter of 50 mm, a 3 mm gap, and a 3 mm strip width (Fig. 1a). Phantom imaging was conducted using the PPWG filled with saline solution, with the metamaterial structure inserted inside (Fig. 1d). The bio-inspired surface coil, positioned outside the waveguide and parallel to the plates, handled both RF transmission and reception. To validate this approach, images were acquired using a standard gradient-echo sequence with the following parameters: TE/TR = 4 ms/100 ms, FOV = 40 × 40 mm², matrix size = 256 × 256, flip angle = 45°, slice thickness = 1 mm, and NEX = 1. For comparison, additional phantom images were obtained using an in-house birdcage coil (length = 5 cm, diameter = 4 cm, 4 rungs). All experiments were performed on a 7T/30 cm Bruker scanner (Bruker BioSpin MRI GmbH, Germany).

Phantom images were acquired both with and without the metamaterial to validate the feasibility of this approach (Fig. 2c). Additional phantom images were obtained using a birdcage coil and the twMRI setup without the metamaterial (Fig. 2c and e, respectively). From the acquired images, SNR values and uniformity profiles were computed along the yellow line indicated in Fig. 2b (see Fig. 3a).

The measured SNR values were: Sandwich-metasurface: 30.4, twMRI (no metamaterial): 24.5, In-house birdcage coil (in-hBC): 32.42. The sandwich-metasurface imaging performance was comparable to that of the in-house birdcage coil—a significant result, given that remote MRI acquisition typically suffers from reduced image quality. While all uniformity profiles exhibited similar patterns, the metamaterial-enhanced setup demonstrated higher signal intensity than the standard twMRI approach. This confirms that even a non-tuned metamaterial provides a notable improvement over conventional remote imaging methods, aligning with prior findings at 3 T [4] and 15.2 T [1].

These experimental results demonstrate that a metamaterial-loaded PPWG filled with saline solution can produce high-SNR images using the tw MRI approach. Our findings show that this method outperforms conventional twMRI techniques at 7 T in a preclinical imaging setting, offering a promising alternative for high-quality remote MRI acquisition.

Metamaterials have demonstrated the ability to enhance the performance of RF coils in clinical MRI [1]. Preclinical MRI also requires RF coils with enhanced performance, and metamaterials provide a promising alternative to achieve this goal [2-3]. We have previously reported the experimental results using a flexible metasurface wrapped around a phantom and bio-inspired surface coil [4]. In this paper, we developed a metasurface and bio-inspired coil as a single unit: a metasurface based on the Hilbert curve [5] and a bio-inspired coil [4] were integrated into a single unit for preclinical MRI applications at 7 Tesla.

The size of the metasurface must correspond to that of the bio-inspired coil and its resonant frequency. We employed the theoretical framework developed by Chen et al. [6] to calculate the frequency for the principal mode of the Hilbert metamaterial: fm=mc/2(2^N+1)a (1) where m represents the harmonic number, c is the speed of light, N is the order and a is the side length. Using eq. (1) with N = 4, m = 1 and a = 3 cm, we obtained the following resonant frequency, . The prototype was constructed using copper sheets (thickness = 35 microm and s = 5.96 x 10^7 S/m) laminated onto a nonconductive board (FR4: e = 4.35 and tan(d) = 0.008, thickness = 1 mm, 6.5 cm long and 3.5 cm wide). The bio-inspired coil was laminated onto one side of the substrate, while the Hilbert fractal curve was laminated to the opposite side. Tuning and matching capacitors (0–15 pF: Voltronics Co., Salisbury, MD, USA) were directly soldered onto the surface, incorporating six parallel ceramic capacitors (American Technical Ceramics, Huntington Station, NY, USA) to achieve a total capacitance of 27 pF, as illustrated in Fig. 1.a). The coil prototype was precisely matched and tuned to 50 Ω at a frequency of 299.471 MHz, corresponding to the proton frequency for 7 T. Notably, the metasurface was not tuned or matched using passive components, as its resonant frequency was sufficiently aligned with the experimental frequency required for these tests. Fig. 1 presents a photograph of both the coil prototype and the Hilbert metasurface. To assess the effectiveness of this new coil design, phantom images were captured using a standard gradient echo sequence, with parameters set to TE/TR = 4.39 ms/200 ms, FOV = 60 mm × 60 mm, matrix size = 128 × 128, flip angle = 30°, slice thickness = 1 mm, and NEX = 1. Additionally, phantom images were obtained using a bio-inspired coil of comparable dimensions without the metamaterial for comparative analysis. All MRI experiments were performed on a 7T/30 cm Bruker imager (Bruker BioSpin MRI, GmbH, Germany).

The resonant frequency of the Hilbert curve, calculated using Eq. (1), aligns closely with results reported by Motovilova and Huang [7], given similar dimensions and orders. It is important to note that in our case, the Hilbert curve metamaterial is not employed as a resonator, and no passive components were used for tuning or matching purposes. Instead, the geometrical characteristics and dimensions of the metamaterial determine the resonant frequency, making it suitable for preclinical applications at 7 Tesla. Figs. 2a) and b) display phantom images acquired with the Hilbert metasurface-integrated surface coil and the coil without the metasurface, both exhibiting excellent image quality. The signal-to-noise ratio (SNR) values were calculated from these images, resulting in SNRmeta+coil = 120 and SNRbiocoil = 80. This indicates an approximate 66% improvement in SNR for the metasurface-integrated coil. Additionally, a comparison of SNR versus depth was conducted experimentally using phantom images for both coil prototypes, as shown in Fig. 2c).

The Hilbert metasurface-integrated surface coil demonstrates a significant performance enhancement over the bio-inspired surface coil, particularly in terms of field uniformity, signal-to-noise ratio (SNR), and spatial selectivity. These improvements are attributed to the metasurface, which enable more efficient manipulation of the RF field distribution. These experimental results are consistent with previous data reported by our group [4], further validating the reliability and reproducibility of the metasurface-based design strategy in high-field MRI applications.

The experimental imaging results indicate that the Hilbert metasurface-integrated surface coil outperforms the bio-inspired surface coil. By incorporating a metamaterial into the surface coil, we can enhance its performance without the need for additional electronic components, thereby facilitating the development of innovative coil designs. Our findings demonstrate that integrating a metasurface can significantly improve coil performance for preclinical applications in high-field environments.

Standard clinical MRI is outperformed by ultra-high-field MRI in terms of image signal-to-noise ratio (SNR). Bringing this level of quality into the clinical setting could enable more accurate pathology detection, earlier diagnosis, improved therapeutic monitoring and a deeper understanding of cardiac disease mechanisms. Our goal is to achieve such image quality in clinical MRI by increasing spatial resolution, which requires a higher SNR. The performance of the receive coil is a critical determinant of SNR. Conventional surface coils often provide insufficient SNR in standard clinical MRI due to their distance from the heart and relatively large size. The design of an intracardiac receive coil that can be deployed within an intravascular catheter and achieve submillimetre resolution is presented in this study.

The aim of this study is to develop a cardiac imaging catheter coil for use in 1.5T MRI scans, either intracavitary or transesophageal (fig1). Receive-only coils were developed and connected to a dedicated interface. To maximise SNR, the loop diameter was set at 2 cm. A decoupling method developed in our lab was used to minimise nearby electronic components and reduce RF-induced heating. Miniaturised parts included 0505-sized non-magnetic capacitors (Passive Plus, USA) and 1 mm coaxial cables (ES-France), selected based on electromagnetic simulations for signal optimisation, using Advanced Design System (ADS, Keysight Technologies, USA) and QucsStudio (RAFIsoft, Germany). To suit intracardiac and intravascular navigation, the coil was made mechanically robust and flexible to withstand cardiac motion and anatomy. It was built on a flexible polyimide PCB and bonded to a nitinol support for shape memory. Nitinol is MRI-compatible, and the adhesive was LOCTITE 4305. The total catheter length was ~100 cm. The shaft used MRI-compatible materials: • Outer shaft: PEBAX 63D, 3.6 mm outer diameter, with embedded MRI-safe LCP (Liquid Crystal Polymer) braid, previously tested. • Inner shaft: PEBAX 72D with embedded MRI-compatible LCP braid. The catheter coil was tested in vitro for MRI performance on a 1.5T scanner (Aera, Siemens Healthineers, Germany). Two experimental setups were evaluated: immersion in a saline-filled bucket (fig2.A), and insertion into a half-heart phantom with saline and 1% agarose (fig2.B). Safety testing used PRFS-based MR thermometry to assess RF-induced heating [1]. Finally, high-resolution ex vivo images were acquired in a sheep heart to validate imaging performance (fig3).

The catheter coil demonstrated excellent performance, with a 13-fold higher SNR within a 2 cm diameter region compared to conventional coils. Similar results were observed in the half-heart phantom filled with agarose gel. The remote decoupling strategy proved highly effective, ensuring efficient isolation of the coil element without physical intervention. The deployment mechanism functioned reliably, enabling precise and repeatable positioning. Ex vivo acquisitions (fig3) also yielded very high SNR values with an isotropic spatial resolution of 500 µm³, approaching the image quality reported in [2] using a non-deployable surface coil of identical diameter. MR thermometry confirmed safe RF performance, with no noticeable temperature increases on or around the coil.

These results demonstrate the feasibility and performance of a flexible miniature catheter coil designed for intracardiac or transesophageal imaging at 1.5T. The high-resolution ex vivo images and high SNR obtained in both phantom and biological tissue confirm the effectiveness of the coil design for localized signal reception in deep anatomical regions. The remote decoupling strategy was effective in significantly reducing the number of electronic components in the imaging zone, thereby minimising the risk of RF heating. This was confirmed by MR thermometry, which showed no measurable temperature increase, indicating that the device can operate safely under MRI conditions. The combination of flexible PCB and nitinol provided both mechanical flexibility and shape memory, allowing the coil to adapt to dynamic cardiac environments while maintaining MRI compatibility. Further improvement of the device may include adding an inflatable balloon surrounding the coil at the tip of the catheter to facilitate coil deployment in constrained regions. Such a system would allow controlled mechanical expansion of the coil, ensuring proper deployment and contact with surrounding tissues, even in tight or stiff anatomical areas. This enhancement could further improve image quality and reproducibility in future in vivo applications.

This work demonstrates the successful development of a compact, remotely deployable catheter coil capable of high-resolution structural imaging at 1.5T. The next steps to finalize the prototype include integrating cable traps to suppress common mode currents, addressing occasional imaging artifacts, and advancing toward in vivo imaging studies.

Simultaneous EEG-fMRI at ultra-high field (e.g. 7T) offers high temporal (sub-millisecond) and spatial (sub-millimeter) resolution, enabling non-invasive exploration of human brain function at the meso-scale. However, EEG-fMRI at 7T poses significant challenges, particularly EEG signal corruption due to the MR environment. One major issue is the ballistocardiogram (BCG) artifact on the EEG signal, mainly caused by heartbeat-induced movement in the static magnetic field. One technique to reduce BCG involves adding artifact detection loops (ADLs) on top the EEG signal detection electrode cap to directly record the BCG artifact from the EEG signal[1]. While this design was successfully implemented on lower magnetic fields, their design has not been optimised for 7T and, owing to the reduced RF wavelength, may cause significant RF field disruption that can affect image quality and subject safety. Using electromagnetic field simulations, this study examines how RF fields are pertubed by ADLs, and how this is altered by ADL size and material type.

Electromagnetic field simulations were performed using a finite-difference time-domain simulation software (Sim4life 8.0,ZMT,Switzerland). A 16-leg high-pass shielded birdcage coil (diameter=305mm, length=210mm) driven in circularly polarized mode was simulated, and tuned and matched using Optenni (Optenni Lab, Finland). This coil was loaded with a digital model of a phantom (SAM phantom,ϵ_R=45.3,α:0.87S/m,ZMT,Switzerland). A model of 4 ADLs was built based on a 7T prototype cap (BrainCap-MR7FLEX, Brain Products GmbH), resulting in wire lengths of approximately 28cm corresponding to the open loop area. These wires were then transformed to obtain lengths corresponding to 50%, 60%, 70%, 80%, 100% and 120% of the open loop area. Simulations for all wire lengths were conducted for carbon wire (CW)(ϵ_R=5,α=62695S/m) and resistive polymer (ϵ_R=5,σ=5.556S/m). Loops were projected onto the surface of the phantom and displaced 7mm away from the surface in both x and y directions to maintain constant spatial relations between the loops and phantom. The cable tree, where the wires from the loops aggregate and exit the RF coil was modelled as carbon wire and maintained constant for all simulations. Figure 1(a) depicts the simulation model. A simulation with no loops (just phantom and RF coil) was run to establish the unperturbed B1+ and SAR values. Simulated B1+ field maps and 10g mass average SAR (SAR10g) maps were normalised to 1W total input power. These were masked to the region inside the phantom, interpolated at 1mm-isotropic resolution, and exported to Matlab (R2024a,The Mathworks,Natick,MA). Maximum Intensity Projection (MIP) B1+ and SAR maps were computed in all 3 directions for each case. From each SAR10g MIP map, the MIP map of the phantom only case was subtracted to obtain a MIP difference map, to understand the full extent of the effects of the loops on SAR10g.

Figure 1(b) presents the B1+ MIP maps at each loop length for carbon wire and resistive polymer loops. In the carbon wire case, interactions with the B1+ field generally increase with increasing loop length, with maxima seen at 70–80% of the original length. In the carbon wire case, interactions with the B1+ field increased with loop length, peaking at 70-80% of the original length, before decreasing again. At these lengths, the wire creating the open loop area was approximately a quarter of the RF wavelength in air at 7T, matching resonance conditions previously reported for EEG cable bundles.[2] In contrast, minimal B1+ field perturbations were observed with higher resistance polymer ADLs, irrespective of loop length. Figure 2 and 3 show the SAR10g MIP and MIP difference maps, indicating similar trends to the B1+ maps. In the carbon wire case, resonance effects at 70-80% loop lengths led to substantial increases in SAR, whereas no such increase was seen in the polymer case due to the resistive properties of this material. As depicted in Figure 4, SAR10g values at these lengths were increased by 21-26% compared to the phantom-only case. For the polymer, SAR10g remained stable across all loop sizes, showing no resonance effects.

Carbon wire ADLs were more susceptible to RF interactions than resistive polymer ADLs. Resonance effects were observed in the carbon wire loops at lengths corresponding to a quarter of the RF wavelength at 7T, consistent with previous findings in EEG cable bundles.[2] The polymer material, with its lower conductivity, exhibited minimal RF interaction, regardless of loop size.

These results suggest that ADLs made from moderately resistive materials (like carbon wire) should avoid resonant lengths to limit RF field interactions. In contrast, highly resistive loops (e.g. polymer) do not exhibit significant RF interference, regardless of size. Optimizing ADLs for EEG artifact suppression while maintaining MRI data quality and safety will require considering these material and loop size effects.

Breast MRI is typically performed in the prone position. Supine breast MRI offers better anatomical alignment with other imaging modalities and surgical planning, but compromised by respiratory motion artifacts. We previously showed that GRICS correction with a respiratory belt improves supine breast MRI [1]. With BraCoil [2], a flexible supine breast coil, accelerometers mounted on its surface outperformed the belt - likely due to better correlation with chest motion [3]. Despite its flexibility, BraCoil may still dampen motion signals. Therefore, direct chest placement of sensors is expected to improve signal fidelity. However, MEMS-based accelerometers require batteries, which introduce susceptibility artifacts and must be positioned away from the tissue of interest. In this study, we assess the MRI compatibility of an optical accelerometer and demonstrate its usability for respiratory motion correction in supine breast MRI.

The evaluated sensor was the optical accelerometer FOSA 3660 (Optoacoustics Ltd, Or Yehuda, Israel). The MRI was Siemens Prisma 3T. Testing included safety, interference, and motion correction performance. Safety tests assessed RF-induced heating and gradient-induced vibrations. i) RF heating: The sensor was installed on a phantom (Fig. 1a). Four optical temperature probes were connected to a Reflex conditioner (Neoptix Canada LP). Temperature data were recorded in 3 phases: pre-sequence (20 min), during an MRI sequence (qTSE, RF Level 1, 15 min), and post-sequence (5 min). To account for potential thermalization, the pre-sequence data were fitted using Newton’s law of cooling. The difference between the predicted temperature and the actual post-sequence temperature was then calculated. ii) Vibrations: The accelerometer was mounted on a foam support in the MRI bore (Fig. 1c). Vibrations perpendicular to its main surface were measured using a laser vibrometer in combination with a custom fixture equipped with a prism. 8 EPI sequences were applied at various frequencies [4]. The average root mean square (RMS) displacement of the accelerometer was compared to that of a plastic block of similar shape and size. Interference testing addressed susceptibility artifacts and RF noise emissions. For the susceptibility artifacts, a GRE sequence was applied (TR = 500 ms and TE = 26.6 ms). The artifacts were compared to those induced by an MRI-compatible accelerometer [5] with a non-magnetic battery PGEB-NM651825-PCB (PowerStream, Toronto, Canada). RF noise emission was assessed using a Siemens diagnostic tool that sampled the RF spectrum received by the MRI body coil over the resonance frequency range of 123.2 ± 0.5 MHz, without RF excitation. To evaluate motion correction performance for supine breast MRI, the sensor was positioned on the chest of a female volunteer. A T2w TSE sequence (60 slices) was run twice: once during forced, hard chest breathing, and once during abdominal breathing. Sensor signals recorded during the hard breathing were low-pass filtered (0.3 Hz) and used for the GRICS motion correction algorithm [6]. Resulting image quality was evaluated using the sharpness index [7].

Fig. 1b shows the results of the RF heating test. The probes started slightly warmer than the MRI room (by 0.1 to 1.2 °C). The estimated temperature increase of the far end of the sensor box (“side”) during the sequence was about 0.1°C and smaller for other parts. Vibration analysis (Fig. 1d) showed a mean RMS displacement of 0.26 ± 0.13 µm, similar to that of the plastic block (0.24 ± 0.06 µm). As shown in Fig. 2, the susceptibility artifact caused by the optical sensor was approximately 4 mm deep, compared to 32 mm artifact from a battery-powered accelerometer. RF spectra acquired with and without the sensor showed no detectable difference, confirming the absence of emitted RF noise. Fig. 3a displays respiratory signals recorded during the first TSE sequence, and Fig. 3b shows the resulting images. The motion correction improved image sharpness by 21.5 ± 13.1% (p < 0.001).

The RF heating tests confirmed that the sensor is safe. Even assuming an initial body temperature (37 °C), it remained well below the 43 °C safety threshold (IEC 60601-1). It exhibited no significant induced vibrations across all tested EPI sequences, reducing the likelihood of patient discomfort, mechanical failure or added noise during acquisition. Minor susceptibility artifacts were observed. Nonetheless, the artifact size was substantially smaller than that of a battery-powered accelerometer and would likely be even smaller under less extreme imaging parameters. Finally, the sensor was successfully applied for the GRICS motion correction for supine breast MRI. Validation on more volunteers and comparisons with other sensors are needed to further assess its performance.

The optical accelerometer is safe for direct chest placement in the MRI environment, introduces minimal artifacts, and can be effectively used for GRICS motion correction.

Gradient-array coils add extra degrees of freedom over conventional gradient systems. They have already enabled simultaneous multi-slice (SMS) excitation without SAR penalty [1], region-of-interest (ROI) gradient focusing for efficiency gains [2], and lower electric-field exposure that lifts peripheral-nerve-stimulation (PNS) limits [3]. Still, developing a practical and scalable gradient power amplifier (GPA) system has been challenging. Conventional gradient amplifiers are large, expensive, and difficult to scale beyond a small number of channels. As the number of channels increases, interactions between coils become more significant, introducing coupling effects that complicate control and require more advanced strategies. Earlier work leaned on feedforward methods that compensated for coupling using known coil models, but feedforward control strongly depends on accurate modeling, making it sensitive to model inaccuracies [4,5]. Feedback control, in contrast, directly addresses coupling effects, though it traditionally relies on costly fluxgate sensors. A low-cost analog controller was also evaluated [6], but its manual tuning becomes difficult as channel count increases, leading us to focus on digital feedback in this work. In the current implementation, we use fluxgate sensors, though the system is designed to allow future use of cheaper current sensors. We present a modular gradient controller using high-bandwidth digital feedback. With GaN transistors and continuous-time delta-sigma ADCs, the system drives coupled coils with stable current control. Our earlier test setups involved loosely connected boards and wiring, making integration difficult and raising concerns about robustness. By adopting the Eurocard format, we consolidate the system into a single enclosure with standard mechanical and electrical interfaces, resulting in a cleaner, more maintainable design suited for practical deployment and further development.

Each controller follows the IEEE 1101.1 Eurocard standard. A 3U sub-rack holds twelve modules, allowing up to 168 channels per rack. Our four-channel module includes an Artix-7 FPGA running real-time PID control. Currents are digitized by a 24-bit, 1.5 MSPS AD4134 ADC, which removes the need for anti-alias filters and keeps latency under 10 µs. Feedback is provided by LEM fluxgate sensors. Amplifiers, still under development, were emulated using external GaN-based PWM setups at 666 kHz. Center-aligned PWM doubles the control bandwidth, helping avoid output filters and reducing complexity. A plastic optical fiber (POF) board links the controller to an AMD Kintex Ultrascale FPGA acting as the MRI spectrometer, which parses Pulseq data and streams gradient waveforms optically. This minimizes EMI and is suitable for future operation inside the scanner room. Validation used four z-elements of a 40-channel custom gradient coil in a double Maxwell configuration. This setup introduces mutual coupling, allowing the controller to be tested in conditions relevant to array systems.

The digital feedback controller successfully drove four mutually coupled gradient channels, maintaining stable and independent currents. A spin-echo MRI scan of a tomato was performed using our system for the readout gradient, while phase encoding and slice selection were handled by the scanner’s gradients. The resulting image and current measurements confirm that the controller operates reliably under coupling. At higher gradient currents, we observed image noise that likely originated from EMI at the Larmor frequency, probably caused by PWM harmonics. Twisted shielded cables and amplifier shielding significantly reduced the noise, confirming its electromagnetic origin. These results underline the importance of EMI mitigation, especially in systems without conventional output filtering.

The controller captures both current and drive signals at high bandwidth, making it possible to estimate the coil and amplifier response from data. The estimated model can be fed forward while feedback corrects the remaining error, allowing predictive or adaptive control. This may let the system work with cheaper sensors instead of fluxgates, without losing stability or tracking.

We built a modular digital feedback controller and tested it on four mutually coupled coils. The system maintained stable currents and produced a clean spin-echo image using the array for readout. The design supports future expansion to larger arrays and offers a base for exploring model-based control and alternative sensor types.

Acquisition of robust Magnetic Resonance Imaging (MRI) data relies on a homogeneous B0 field, but magnetic susceptibility differences in vivo can create B0 distortions that lead to artifacts and signal dropout [1–3]. Multi-coil (MC) shimming systems use an array of individually-driven generic coils to homogenize the B0 field, and have been shown to outperform low-order spherical harmonic (SH) shimming in the brain [4–8]. Most MC setups are designed as temporary inserts for existing scanners, with high-quality basis maps [5, 7–9] (maps of the field produced per shim setting for each coil) acquired by a calibration process requiring multiple hours of scan and analysis time. Shim fields are calculated as a combination of calibrated basis maps for subsequent experiments. For an applied shim field to match the calculated field, it is therefore essential to either position the hardware in precisely the same location as in calibration or have knowledge of the exact position relative to the calibration scenario [5]. Degradations in B0 field control are observed even for millimeter-scale displacements (Figure 1). Our purpose is to reliably detect misplaced hardware using only a field map acquisition, ensuring optimal performance of MC inserts.

A hardware self-registration algorithm [10] in MATLAB (Mathworks, Inc., Natick, MA) co-registers two 3D field maps: an Expected Field, which is a field map acquired at the Day 0 (i.e. calibration) position, and a Measured Field, which is the same field shape measured at Day N (Figure 2). The Expected Field is mathematically created by a combination of the calibration maps, and the Measured Field is measured after the hardware is re-placed in the scanner. The algorithm calculates the x, y, and z-translation and z-rotation needed to align the two fields in an eroded ROI through co-registration. Secondary rotations about the x- and y-axes can be considered small in practice and were, thus, not considered in this proof-of-concept. Three Expected Field shapes were tested: the shape of a selected single MC element as a baseline, a Four Lobe field generated by only one ring of coils, and a GA Field optimized by a genetic algorithm (Figure 4). The latter two were selected for further validation. A MC array comprising 6 rows of 8 coils (diameter 70 mm, 100 turns) was arranged on a cylindrical former (OD 20.32 cm) to produce B0 field distributions (Figure 3). Monte Carlo simulations of 5,000 recoveries of rigid transformations, were used to calculate the average norm error of localization accuracy at ten SNR levels. Simulations were done for Simulated Basis Maps – the maps generated from Biot-Savart simulations of the hardware – as well as the calibrated Experimental Basis Maps for the hardware. The self-registration procedure was tested in the scanner by physically shifting the MC hardware a known distance. Scanner validation was performed at seven SNR levels for one hardware position, and then at one SNR level (19±1) for ten random hardware positions.

The established framework enabled accurate hardware localization through application of unique field shapes (Figure 4). Localization accuracy with well below 1 mm and 1 degree errors was achieved irrespective of the applied test field shape with SNR of at least 10. In scanner validation, both Expected Fields had an average translation error below 1 mm for SNR levels above 4.6, and an average rotation error below 1 degree for SNR levels above 2.8 and 4.6 for the GA and Four Lobe fields, respectively. For all SNR levels, the Experimental Basis Maps had higher errors than Simulated Basis Maps. Both Expected Fields were able to recover all 10 additional controlled hardware shifts within a norm translation error of under 0.44 mm, with an average error of 0.30 mm and 0.20 mm for the Four Lobe and GA Fields, respectively. All rotation errors were recovered within 0.17 degrees, with average errors of 0.13 and 0.11 for the Four Lobe and GA Fields.

A field shape optimized with a genetic algorithm allows for an average localization accuracy of 0.2 mm and 0.11 degrees from a sub-1-minute B0 mapping experiment (SNR 20), allowing for improvement of B0 control and shimming potential of hardware inserts. Simulated, Experimental, and Measured experiments showed excellent agreement across SNR levels. Both Expected Fields achieved error below 0.5 mm and 0.5 degrees for the 10 transformations – a reasonable threshold for success given the degradation curve in Figure 1 – but the GA Field had an overall lower error. In the future, we plan to include x- and y-rotations, which cannot be accounted for by rigid transformations alone and require qualitative updates of the basis shapes, and to test the self-registration method in vivo.

Here we have presented a method for field map-based hardware self-registration for MC inserts. This addresses the requirement of precise repositioning of MC hardware inserts, allowing for excellent shim capabilities of misplaced hardware.

Advancements in Very-Low Field (VLF) MRI systems have reduced the complexity of the installation, enabling the mobility of MRI equipments. However, this comes at the cost of lower Signal-to-Noise Ratio (SNR). At 50 mT and without a Faraday cage, the human body acts as an antenna coupling external electromagnetic interference (EMI) into the imaging region [2]. Recent low-field MR literature reported noise reduction via coil optimization, EMI detection and subtraction using multiple external coils, and passive methods like grounding the patient via conductive cloths or pads [1]. Software-based deep-learning approaches have also been explored to suppress EMI [3]. In this study, we examined the impact of patient grounding on EMI coupling and identify some of the factors influencing its effectiveness in VLF MRI.

We hypothesized that patient grounding can be optimized by understanding its dependencies. A custom cylindrical phantom was constructed to replicate human electrical properties [4]. The phantom consisted of a 160-cm long, 12-cm diameter polyvinyl chloride (PVC) tube, filled with water or saline (9 g/L [5]). Two non-ferromagnetic brass screws were inserted at different positions along the tube, penetrating the PVC wall to make contact with the saline. The phantom was placed in a VLF MRI system. Noise measurements were performed using a solenoid transmit/receive coil and a Magritek Kea2 spectrometer with its «MonitorNoise» sequence. We assessed grounding effectiveness as a function of: conductivity of the medium, distance from the imaging region (grounding position) and the presence/absence of direct electrical contact. Noise was repeatedly recorded during 60 seconds in 50-ms time windows (4 averages). Signals were concatenated, and the standard deviation (SD) was computed for comparison. SDs were first compared w/o grounding using grouding point 1 with both water and saline to isolate conductivity effects. Two main experimental conditions were then tested using saline only (Figure 1): 1.Grounding position: screws 20 cm apart, both 50 cm from the imaging region: Config. 1: grouding point 1 only, Config. 2: grouding point 2 only, Config. 3: both grouding points simultaneously; 2.Insulation effects: a 110x20 cm conductive belt (σ=3,3.10^4 S⁄m) was wrapped around the tube and grounded to evaluate the effect of capacitive coupling (PVC as insulation): Config. A: position A, Config. B: position B, Config. C: spiral wrap (maximized contact area).

Noise SDs were computed from the real part of the signal. In the conductivity comparison, the saline-filled tube showed higher baseline noise than water, but grounding significantly dropped it, yielding a lower final level (Figure 2). As shown in Figure 3 (a), conductive grounding reduced the noise SD: grouding point 1 reduced SD by 93.7%, versus 36.3% for grouding point 2. Simultaneous grounding through both grouding points yielded a 98.3% reduction. Capacitive coupling results (Figure 3 (b)) showed 79.4% reduction with the belt in position A, 95.2% in B, and 98.5% when spirally wrapped (C).

The comparison of water vs saline filling confirms that when the entire volume is conductive, grounding is more effective, likely due to improved current flow and a more uniform electric potential distribution within the phantom. Both conductive and capacitive grounding can significantly reduce EMI noise in an ULF MRI setup. Using conductive grounding points resulted in the highest noise suppression, with simultaneous grounding at multiple points providing the greatest improvement. Capacitive coupling via belts was most effective when spirally wrapped. Direct contact grounding showed better efficiency when the grounding points were closer to the imaging region, likely due to the interception of external EMI effects from affecting the remaining exposed part. Conversely, capacitive coupling to ground showed better performance at longer distance from the bore and with larger contact areas.

These findings suggest that both grounding strategies can be applied to reduce noise in practical MRI environments, with capacitive methods showing promise for future clinical use due to better patient comfort. Future work will map location, contact, and area dependencies for both methods to maximize noise reduction while maintaining ease of implementation in clinical or portable systems. Acknowledgements This work received support from the french government under the France 2030 investment plan and French “Investissements d’Avenir” programme, as part of the Initiative d’Excellence d’Aix-Marseille Université, A*MIDEX : AMX-2023-CI-01, AMX-23-CPJ-10 and AMX-23-EQ-FO-009, as well as ANRT CIFRE 2024/1145.

Magnetic resonance imaging of the lung is a challenging imaging modality given the low proton density, and short T2* relaxation times. The short T2* can be compensated by using ultra-short echo time (UTE) sequences [1]. A UTE sequence with a 3D center-out radial trajectories has been developed to capture 4D (3D+time) images of the lung [2]. One of the challenges of this sequence is the ultra-short echo time and the radial center-out trajectory making this sequence highly demanding for gradient performance. These techniques often rely on non cartesian trajectories requiring a more precise gradient control where any difference between command and output can result in gradient delay artefacts (Fig.1a) and in a lack of accuracy in the derivation of the navigator. Gradient distortions are caused by non-linearities in amplifiers and variations in coil manufacturing. To tackle these issues, current advanced methods use MRI sequences along with calibrated passive phantoms [3]. Alternatively, we suggest using high-speed magnetic field cameras (MFCs) paired with algorithms that can determine the necessary adjustments from measurements to modify MR sequences and correct gradient distortions. MFCs are arrays of magnetic sensors that allow understanding complex magnetic fields by mapping the complete magnetic field vectors over a volume. MRI poses a significant challenge for MFCs due to the high-speed magnetic fields and large volumes involved. As a result, the complexity of the reading electronics increases with higher sampling rates and a larger number of sensors. Commercial MFCs [4] that utilize NMR sensors are designed to map B0 but do not provide complete field vector information. Currently, Hall-effect sensors are the most suitable type of magnetometer for MR gradients.

The magnetic camera used in this work have been presented in [5]. The MFC has an array of 7x7 integrated 3D Hall sensors (TMAG5273) (Fig.1b), covering a total area of 31.36 cm². These sensors are set to their maximum conversion rate and dynamic range of ± 266 mT. In this setup, the sensors demonstrate an rms noise of 140 µT, translating to a magnetic field resolution of 1.4 μT/√Hz. All sensors in the array are read synchronously in parallel at 7 kHz. Further details about our MFC can be found in [5]. It was then placed at the entrance to a Siemens Magnetom Free.Max (0.55 T) tunnel to characterize the gradients waveforms of a custom UTE sequence with specific a trajectory designed to enable concurrent field monitoring, sampling 25 temporal points per spoke via a field camera.

Fig. 2 illustrates both the predicted gradient trajectories and the signals captured by the 49 sensors of our MFC. Given that our sensor's range is limited to 266 mT, the gradients along the B0 axis (Z-axis) are masked for our camera. Consequently, the subsequent figures focus solely on the X and Y axes. Fig. 3 presents a direct comparison between the results obtained by the central sensor of the camera and the planned gradient trajectories for the X and Y axes. To facilitate comparison, all signals are normalized between -1 and 1. This figure suggests that the measurements taken with the MFC proceeded as intended, as the measured shapes closely resemble the expected ones. The blue areas, which do not correlate with an orange area, correspond to the rewinding gradients applied by the scanner. Additionally, due to the MFC not being perfectly aligned within the MRI, each axis of the MFC detected contributions from gradients other than the “aligned” one. This misalignment is particularly noticeable in the center of the top left graph in Fig. 3. Finally, Fig. 4 provides a detailed view of the measurement of a single trapezoid (comprising 25 points acquired by the MFC) on the X-axis, in relation to the expected shape.

Precisely measuring gradient errors could facilitate the implementation of impulse response function-based correction algorithms [6]. The results obtained with the MFC are promising in this regard but require further refinement of the experimental protocol and electronic development before they can be fully used for correcting sequences. For instance, enhancing the protocol will necessitate improved synchronization between the MFC and the MRI console. Precise positioning of the camera will also ensure cleaner signals. Enhancing the electronics to improve the camera's speed and precision will enable the collection of improved data. Indeed, the sequence used in this work had to be slightly slowed down to be measurable with the MFC. In the future, it will be essential to characterize the full-speed sequences directly.

We successfully demonstrated in this work the ability of our MFC to measure gradients from a UTE sequence with a 3D center-out radial trajectory. The results obtained pave the way for the adjustment of such a sequence to improve the quality of the 4D images of the lung that it enables.

Magnetic Resonance-guided radiotherapy (MRgRT) has emerged as a significant advancement in the field of radiation oncology. The Elekta Unity MR-Linac represents a fusion of high-field diagnostic MRI with a linear accelerator, enabling real-time imaging during treatment and the possibility of daily Adaptive Radiotherapy (ART). This integration allows personalized adaptation of the treatment plan to anatomical changes, improving target precision, reducing exposure to healthy tissues, and enabling safe dose escalation. This study explores the impact of Elekta Unity on treatment workflow, dosimetry in magnetic fields, and the application of real-time motion management techniques.

The Elekta Unity MR-Linac combines 1.5 Tesla MRI with a 7 MV linear accelerator. Its capability to perform daily ART includes two main workflows: "Adapt to Position" (ATP) and "Adapt to Shape" (ATS). ATP compensates for inter-fractional translational setup errors, while ATS accounts for complex changes including rotations and deformations. Treatment planning involves Magnetic Resonance Imaging (MRI), Computed Tomography (CT) for electron density mapping, and synthetic CT generation. The Particle Transport Algorithm (PTA) within the treatment planning system (TPS) models the dosimetric perturbations induced by the magnetic field on secondary electrons. To counteract these effects, optimized beam configurations and IMRT techniques with multiple fields are applied. The study further investigates the Electron Return Effect (ERE) at tissue-air interfaces, the asymmetric dose distribution, and the implications on buildup regions. Comprehensive Motion Management (CMM) is employed to monitor intra-fractional target motion in real time, using non-invasive, non-surrogate-based tracking. The CMM system ensures beam delivery only when the target is within a predefined gating envelope.

Elekta Unity significantly improves soft-tissue visualization, enabling more precise identification of tumor boundaries compared to CBCT. MRI sequences such as T2-weighted, T1-weighted, and FLAIR can be acquired in under four minutes, streamlining the adaptive workflow. Dosimetric analysis revealed that magnetic fields induce a reduction in the electron path radius, modifying dose distribution patterns, especially at heterogeneities like tissue-air interfaces. Dose perturbations include an increased surface dose and asymmetric penumbra. These effects are mitigated by employing multiple beam directions in IMRT planning. Adaptive planning via weight and shape optimization (Method E) allows for complete fluence re-optimization and re-segmentation, ensuring precise dose delivery despite anatomical variations. CMM enables the detection of target movements such as respiratory motion, sudden shifts, or patient non-cooperation. Clinical applications have shown its effectiveness in various treatment scenarios including SBRT for liver and lung metastases, prostate cancer, and non-cooperative patients. The system automatically interrupts radiation delivery when the target deviates from the planned position, resuming treatment once alignment is restored.

MR-guided ART represents a paradigm shift from static to dynamic treatment approaches. Daily plan adaptation allows for margin reduction and dose escalation with reduced toxicity. However, integrating MRI into the radiotherapy workflow introduces new complexities, including longer session times and susceptibility to intra-fraction motion. The magnetic field's impact on electron trajectories must be considered during planning, especially near inhomogeneities. Elekta Unity addresses these challenges by combining real-time imaging, accurate dose calculation via PTA, and motion management through CMM. The device demonstrates how MRgRT can be safely and effectively integrated into routine clinical practice, offering high precision even in complex anatomical regions.

Elekta Unity enables a new standard in personalized, adaptive radiotherapy through the integration of high-quality MRI with linear accelerator technology. The system improves tumor visualization, compensates for anatomical changes, and mitigates motion-related uncertainties with real-time monitoring. Although the adaptive workflow is more time-consuming, the benefits in terms of precision, safety, and patient tolerance are significant. Future developments in automation and motion prediction will further streamline MRgRT, consolidating its role in advanced oncologic treatments.

We present the development of an instrumental and software solution for the acquisition and denoising of electromyography (EMG) signals during functional MRI (fMRI) data recordings. This solution is based on a dedicated acquisition channel optimized to isolate the considerable inductive noise associated with the commutation of magnetic field gradients. It is completed by a denoising tool for evaluating and subtracting this inductive artifact.

Our solution is to design a DAQ channel perfectly synchronized with the MRI system. This DAQ channel includes BIOPAC MRI-compatible electrodes and cables, two instrumentation amplifiers and a National Instruments NI-CompactDAQ 9178 box fitted with NI-9215 (16-bit analog acquisition) and NI-9401 (digital I/O & synchronization) modules. We have checked that the hardware components do not produce any RF artifacts on the MRI images. Acquisition is controlled by a dedicated software developed under the LabVIEW® environment using the DAQmx driver. This software ensures hardware retriggering on each TR, and the generation for each of these volumes (TR) of its own acquisition clock, enabling the acquisition of a train of samples for the duration of the TR. Each signal sample is thus acquired at exactly the same time as the corresponding sample from the previous TR. Average noise is thus calculated on the basis of perfectly synchronous trains (globally over the entire run or over a sliding time window). The average noise is then subtracted from the initial signal, and the resulting trains are concatenated to reconstruct the signal. Resynchronization at each TR guarantees the absence of phase-shift noise (hard to filter) between all these TRs.

The quality and relevance of the EMG signals was demonstrated by implementing a Go/NoGo task using response buttons equipped with isometric FSR pressure sensors, enabling the detection of very low pressures below the button release threshold. We showed the evidence of the correlation between the FSR sensor signals and the EMG signal, even in the absence of button switching.

The acquisition of EMG signals is very informative. For example, these signals can be used to accurately determine the onset of a motor response, before it is translated into movement or whether the response is inhibited and therefore undetectable using conventional tools (force transducers, response buttons, etc.). This methods allowed to implement new features in fMRI analysis in subtile motor tasks

Our solution solves the problem of cleaning up the inductive artifact by developing a specific hardware chain that isolates the induction noise from its source. The results are highly promising and open up interesting prospects. The prospects that are opening up concern both the equipment, with optimized amplifiers or electrodes adapted to the muscles concerned, and the acquisition methods themselves: we could envisage real-time acquisition/de-noising using, for example, an on-board system that would perform on-line cleaning, as exists in audio, for example.

Lung MRI has been suggested as useful for nodule characterization and lymph node metastasis diagnosis, and DWI is an integral part of MRI for lung MRI, and apparent diffusion coefficient (ADC) maps derived from DWI also useful in these settings1-3. However, a major disadvantage of DWIs is that it is considerably prone to artifacts, particularly susceptibility artifacts at tissue interfaces and image blurring due to image distortion. Reverse encoding direction (RDC) techniques with different approaches have been suggested as useful for reducing distortion artifact and improving image quality and diagnostic performance on DWI in not only neuro, but also head and neck or prostatic MRI4, 5. Currently, no major reports are not assessed the capability of DWI with RDC technique (RDC DWI) for improving image quality and influence on lymph node metastasis diagnosis in lung cancer patients. We hypothesized that DWI with RDC technique (RDC DWI) was more useful than conventional DWI (cDWI) for improving image quality and differentiation of metastatic from non-metastatic lymph nodes on lung DWI in patients with non-small cell lung cancer (NSCLC). The purpose of this study was to compare capabilities for image quality improvement and lymph node metastasis differentiation among RDC DWI, cDWI and FDG-PET/CT in NSCLC patients.

40 pathologically NSCLC patients underwent STIR FASE imaging, RDC DWI and cDWI at 1.5T system (Vantage Orian, Canon Medical Systems), FDG-PET/CT at two PET/CT systems (uMRI550: United Imaging, Shanghai, China; or Biograph mCT: Siemens Healthneers, Erlangen, Germany) with same protocols, transbronchial or mediastinal biopsies, surgical treatment, pathological examinations and follow-up examinations. RDC DWI and cDWI were obtained by spin-echo type echo-planar imaging (SE-EPI) sequence by same parameters (TR 3350ms/TE 59ms, b vale 0 and 1000s/mm2, 6 number of excitation [NEX], FOV 300450 mm, 144128 acquisition matrix, 288432 reconstruction matrix, section thickness 6mm, slice gap -1mm, voxel size 1.01.06.0 mm3) with and without RDC technique. According to pathological examination results, 69 metastatic nodes and 69 out of 354 non-metastatic nodes, which were randomly computationally selected, were assessed in this study. For quantitative image quality evaluation, absolute distortion ratios (ADRs) between each DWI and STIR images and signal-to-noise ratio (SNRs) of each selected lymph node were determined by region-of -interest (ROI) measurements. On qualitative image quality assessments, overall image quality, artifact and lesion conspicuity of each lesion and lymph node were assessed by 5-point visual scoring systems. For comparison of diagnostic performance for lymph node metastasis, apparent diffusion coefficients (ADCs) on both DWIs and maximum values of standard uptake value (SUVmaxs) of all metastatic and non-metastatic nodes were also determined by ROI measurements. To compare each quantitative image quality index between RDC DWI and cDWI, paired t-test was performed. To compare all qualitative image quality indexes between two methods, Wilcoxon's signed rank tests were performed. Then, Student’s t-tests were compared to determine differences of each ADC and SUVmax between metastatic and non-metastatic lymph nodes. To compare diagnostic performance among RDC DWI, cDWI and FDG-PET/CT, ROC analyses were performed. Finally, sensitivity (SE), specificity (SP) and accuracy (AC) were compared among all methods by McNemar’s test.

Representative cases are shown in Figure 1. Figure 2 shows compared results of all quantitative indexes between two DWIs. ADR of RDC DWI was significantly smaller than that of cDWI (p=0.003). Figure 3 demonstrates compared result of each qualitative index between two DWIs. Overall image quality and artifact of RDC DWI were significantly improved as compared with those of cDWI (p<0.0001). When compared results of each ADC and SUVmax between metastatic and non-metastatic lymph nodes, there were significant differences of both ADCs and SUV max between metastatic and non-metastatic lymph nodes (p<0.0001). Figure 4 demonstrates results of ROC analysis for diagnosis of metastatic lymph node. Area under the curve (AUC) of RDC DWI was significantly larger than that of cDWI and FDG-PET/CT (p<0.05). When applied each threshold value, SP (100 [69/69] %) and AC (89.1 [123/138] %) of RDC DWI were significantly higher than those of cDWI (SP: 84.1 [58/69] %, p=0.004; AC: 79.7 [110/138] %, p=0.0002) and FDG-PET/CT (SP: 81.2 [56/69] %, p=0.0002; AC: 81.2 [112/138] %, p=0.001).

RDC technique has better potentials for improving distortion, image quality and diagnosis of lymph node metastasis as compared with conventional DWI and FDG-PET/CT in NSCLC patients.

Magnetic Resonance Elastography (MRE) is a non-invasive imaging technique used to quantify the mechanical properties of tissues. This method relies on generating mechanical waves in soft tissues using an actuator synchronized with a motion-encoding gradient (MEG) sequence [1]. The emergence of new actuator designs requires precise characterization of their mechanical vibrations, both outside and inside the MRI environment, to ensure the quality and reliability of MRE-derived mechanical properties measurement [2]. Laser vibrometry has become one of the preferred solutions for measuring vibrations within the MRI environment, thanks to its electromagnetic immunity and non-contact technology [3, 4]. In this study, we characterize the vibrations generated by a commercially available pneumatic passive driver (Resoundant, Inc., United States) both inside and outside the MRI room using laser vibrometry, demonstrating the feasibility of in situ, MRI-compatible vibration measurements.

Vibration measurements, perpendicular to the surface of the passive driver, were performed using a PDV-100 laser vibrometer (Polytec, Germany). For each acquisition, a velocity profile was recorded over 3 seconds at a 48 kHz sampling rate. Thirteen mirrors were distributed across the surface of the passive driver (Fig. 1a). The driver was mounted on a custom plastic measuring bench [5] designed to minimize the transmission of mechanical vibrations (Fig. 1b, c). Vibrations were generated at 20% maximum amplitude capacity using different MRE sequences: i) clinical gradient echo sequence at 60 Hz ii) SE-EPI based 3D MRE sequence at 80 and 100 Hz in a 1.5T MR scanner. For vibration measurements outside the MRI scanner room, the laser vibrometer was aligned perpendicular to each measurement point (Fig. 1b). Velocity profiles were acquired at all points for each MRI sequence. For vibration measurements inside the MRI scanner room, the passive driver was placed at the entry of the MRI tunnel with its main surface parallel to the static magnetic field (B0). Vertical vibrations were measured by redirecting the laser beam through a prism (Fig. 1c). Velocity profiles were acquired only at the central point due to setup limitations. Data were processed using MATLAB (Mathworks, 2024), the velocity signals were bandpass filtered (cutoff frequencies: 5 Hz – 500 Hz), and the displacement profiles and spectra were computed using frequency domain integration [6]. The root mean square (RMS) displacement, fundamental frequency and harmonics together with their associated amplitudes were calculated.

Fig. 2 shows an example of the measurements performed outside the MRI at the central point for all tested sequences. The fundamental frequency matched the input frequency across all points and sequences, with significant harmonics detected up to the third order. The displacement amplitude decreased as the frequency increased. Fig. 3 shows the distribution of the displacement measured outside the MRI along the center line across the X-axis of the device. Consistent with the driver’s design, a bell-shaped profile is consistently observed for all applied sequences. Concerning the measurements performed inside the MRI, Fig. 4 shows the vibrations of the central point compared to the ones measured outside the MRI for the clinical sequence at 60Hz. On average, RMS displacement inside the MRI was 7 ± 3 % higher than outside across all sequences (Fig. 4c). This difference is within the margin of measurement uncertainty inside the MRI, estimated at ±12%(studies not shown here). The fundamental frequency observed matched the input for all sequences, consistent with outside-MRI measurements.

Frequency analysis confirmed that the driver vibrates at the command frequency, both inside and outside the MRI environment. However, as higher frequencies result in lower vibration amplitudes, appropriate compensation must be applied to maintain the same wave propagation. The similarity between displacement spectra inside and outside the MRI indicates that gradient-induced mechanical vibrations potentially transmitted through the pneumatic tube do not affect driver vibrations within the frequency range used for mechanical property characterization. Frequency and displacement measurements inside the MRI were within the uncertainty limits when compared to the outside MRI results, validating the in situ MRI characterization approach. The implementation of a scanning laser vibrometry [7] could reduce characterization time. Improvements to the current setup could enable spatial vibrations mapping inside the MRI

The Resoundant driver demonstrates accurate frequency reproduction and repeatable displacement, inside and outside the MRI environment. The developed setup demonstrates a proof of concept for in situ characterization of MRI elastography systems using laser vibrometry, enabling benchmarking closer to real conditions between devices and risk assessment of vibration effects on patient tissue.

Magnetic Resonance Imaging (MRI) continues to evolve rapidly, driven by innovations in deep learning, quantitative mapping, and hybrid imaging approaches. Recent advances have expanded MRI’s capabilities beyond conventional anatomy imaging to include quantitative tissue profiling [1], super-resolution reconstruction [2], and even real-time therapeutic guidance [3]. As MRI systems grow increasingly complex, there is a growing need for more versatile, anatomically realistic phantoms that replicate both structural detail and tissue-specific contrast behavior to support the testing, development, and teaching of these emerging technologies [4,5]. While many existing phantoms focus on replicating quantitative relaxation times, they remain overly simplistic in structure, lacking the anatomical complexity and clinically relevant relative contrast seen in vivo [4,6], particularly in musculoskeletal imaging, where no anatomically accurate MRI phantom of the knee currently exists. To address this gap, we present the technical development and evaluation of 3D printed sectional (2D) anatomically realistic head and knee phantoms, designed to achieve both qualitative contrast realism and quantitative relaxation times which are critical for MRI technologies development testing and validation.

Our phantom development followed four main steps: (1) manual segmentation of MRI images, (2) CAD-based modeling, (3) formulation of tissue-mimicking materials (TMMs), and (4) 3D printing, filling, and MRI validation (Fig. 1). TMM compositions were selected based on literature [7–11] to replicate contrast behavior seen in T1- and T2-weighted imaging. MRI slices (sagittal, axial, coronal) of the head and knee were sourced from our institutional image library and manually delineated by an expert radiologist. Segmentations were exported as .svg files and modeled into hollow modules using PTC-Creo. CAD models were scaled 1:1 of average size of the specific human anatomy simulated, exported as .stl, and sliced in BambuStudio. Printing was done using 1.75 mm PETG filament on an FDM printer (parameters in Fig. 2a). Each TMM was poured into its corresponding compartment in the 3D printed hollow modules using a beaker or syringe, sealed with film, and stored at 4 °C. Imaging was performed on a 1.5 T clinical scanner (Ingenia Evolution, Philips Healthcare) with a Head-Neck coil. Phantoms were scanned using T1w, T2w and Rhow sequences to evaluate contrast, resolution, and artefact behavior (acquisition parameters in Fig 2b).

Figures 3 and 4 present the qualitative MRI results of the head and knee phantoms respectively. Figure 3a shows a reference axial MRI image of the human knee, while Figures 3b-d display the corresponding phantom images acquired using the protocols outlined in Figure 2, depicting comparable anatomical regions and contrast realism. Figure 4a presents a sagittal MRI image of the human head, included for comparison with phantom images in Figures 4b-d, which illustrate how variations in acquisition parameters, specifically pixel size affect spatial resolution. Figure 4d highlights the anatomical fidelity of the sagittal phantom, with several key structures readily identifiable.

We have developed high-fidelity, anatomically realistic MRI phantoms of the head and knee, achieving relative contrast consistent with clinical expectations in T1-weighted, T2-weighted, and proton density-weighted sequences. By comparing phantom images to human MRI references, we qualitatively confirmed their ability to reproduce tissue-specific contrast patterns and recognizable anatomical features. Beyond visual validation, the phantoms were used to demonstrate a practical use case in image quality assessment. Varying acquisition parameters, particularly pixel size, revealed visible differences in spatial resolution and artefact expression. For example, truncation artefacts (black arrowhead, Figure 4b) were more pronounced in lower-resolution images, while anatomical detail (white arrowhead, Figure 4c) improved with higher matrix acquisition. These underscore the phantom’s utility in sequence optimization and technical benchmarking. The ability to identify multiple anatomical structures further supports the phantom’s application in education and training, providing a reusable, ethically neutral platform for hands-on MRI instruction. While this work was able to achieve anatomical and contrast realism, quantitative relaxometric accuracy (T1 and T2 mapping) remains an ongoing objective. Future efforts will incorporate these parameters to extend the phantom’s application to quantitative imaging research, sequence standardization, and the generation of machine learning-ready datasets.

This work presents anatomically and contrast-representative MRI phantoms of the head and knee, demonstrating their value in imaging protocol evaluation and radiographic training. These models lay the foundation for future development of fully quantitative, multi-purpose MRI phantoms.

Fluid Attenuated Inversion Recovery (FLAIR) is a widely used sequence for clinical diagnosis [1]. The construction of synthetic FLAIR (synFLAIR) images from quantitative MRI (qMRI) data may suffer from CSF partial volume (PV) effects [2,3], yielding false positives. To overcome this issue, several approaches such as deep-learning-based techniques [4,5] have been proposed. Here, a numerical algorithm for synFLAIR image calculation from T1 and T2 maps with suppression of CSF PV and motion artefacts is described.

Data were acquired on a 3T scanner for 5 healthy subjects (29-79y) and 20 patients with essential hypertension (50-70y). For each subject, acquisition comprised: T1 mapping: variable flip angle method, FoV=256mm x 224mm x 160mm, 1mm isotropic resolution, FA=4/24°, TR/TE=16.4/6.7ms, including B1 and B0 mapping [6]; T2 mapping: five turbo spin echo data sets, FoV=256mm x 176mm x 125mm, resolution 1mm x 1mm x 2.5mm, TR=10s, TE={17,86,103,120,188}ms. T1 and T2 calculation was performed according to [7] and [8], respectively. For synFLAIR image calculation, the following parameters were assumed at 3T: T1_CSF=4.5s [9], T2_CSF=2s [10], maximum tissue T2: T2_max=90ms [11,12]. The algorithm comprises the following five steps: (A) Processing of T1 map: A mask of pure CSF is derived from T1>2/3*T1_CSF and subdivided into inner (ventricles) and outer (sulci) CSF (Fig. 1a). CSF PVs are detected according to [13], deriving an enhancement parameter (EP) from T1 gradients (Fig. 1b). A parameter V is obtained by fitting T1 versus EP, highlighting areas of potential CSF PVs (Fig. 1c). These areas are added to the outer CSF via mask growing, resulting in the final mask CSF1 (Fig. 1d). Please note that CSF1 comprises both pixels with pure CSF and with CSF PVs. Tissue R1 (R1_tiss) is derived as 1/T1 outside of CSF and extrapolated across the CSF areas indicated in CSF1. The CSF PV in the T1 map (pv1, see Fig. 1e), is then derived by assuming approximately R1=pv1*R1_CSF+(1-pv1)*R1_tiss. (B) Processing of T2 map: T2 maps are co-registered to T1 maps. A T2-based mask CSF2 is calculated by adding to CSF1 inside a surrounding 2-pixel-layer connected areas with T2>T2_max. R2_tiss and the CSF PV in the T2 map (pv2, see Fig. 1f) are then derived in the same way as for T1. An effective T2 is calculated: T2_eff=T2 (for pv2=0) and T2_eff=1/R2_tiss (for pv2>0). (C) Creation of background image (BI) from T1 map: A pseudo proton density (PD) map is derived from the T1 map according to [9] and a (purely T1-based) BI is calculated as BI=(1-pv1)*PD2 (Fig. 2a). Here, the first factor suppresses CSF, the second factor introduces PD-weighting and an additional pseudo-T2-weighting across white (WM) and gray (GM) matter since the quotient T2(GM)/T2(WM) matches roughly PD(GM)/PD(WM) [11,12,14]. (D) Creation of synFLAIR images: synFLAIR=BI*f(T2_eff) where f provides a T2-based signal enhancement and is given by: f=1 (for T2_eff=T2_max). For TE, 200ms is chosen. (E) Motion correction: To suppress false positives due to motion artefacts in the T2 map, T2-based enhancement is only allowed for areas that also have elevated T1 values (average inner T1 must be at least 1% larger than the surrounding T1).

Figure 2 shows for a healthy subject BI (Fig. 2a), synFLAIR (Fig. 2b) and conventional 2D FLAIR (Fig. 2c) and 3D FLAIR (Fig. 2d). synFLAIR does not show any false positives due to CSF PVs. Figure 3 shows for a patient with WM lesions (79y, arterial hypertension) four slices from synFLAIR (Fig. 3a) and from a conventional 3D FLAIR (Fig. 3b). synFLAIR highlights identical areas and is again free from false positives due to CSF PVs. Figure 4 shows the effects of motion correction: motion artefacts in the T2 map (Fig. 4a) result in false positives in the uncorrected synFLAIR (Fig. 4c). As the respective areas are not salient in the T1 map (Fig. 4b), erroneous enhancement is suppressed in the corrected synFLAIR (Fig. 4d).

The presented algorithm for synFLAIR image generation from quantitative T1 and T2 maps is free from false positives due to CSF PV effects. It also includes the suppression of motion artefacts in the T2 map which are prone to cause erroneous hyper-intensity in synFLAIR. The synthetic FLAIR images presented here enhance the same areas as conventional FLAIR data. A potential improvement would be an additional correction for motion artefacts in the T1 map as these may reduce the quality of the CSF1 mask and hamper CSF suppression.

The calculation of synthetic FLAIR images from T1 and T2 maps may help to save the scanning time required for conventional FLAIR acquisition in studies employing qMRI techniques. Please note that the potential of qMRI is not limited to the construction of synthetic FLAIR images, but extends to any type of weighting, offering great potential for saving scanner time and improving diagnosis.

Cerebrovascular reactivity (CVR) measures capacity of blood vessels to respond to demand, which is reduced in several vascular diseases [1]. CVR can be measured using blood oxygen level dependent (BOLD) MRI, typically during a block-design paradigm alternating between medical air and air with a small proportion of CO2, to induce hypercapnia. As an in vivo measure, assessing the effect of potential confounders on CVR can be challenging. Previously, different processing approaches have been assessed using simulations [2]. However, such simulations and most CVR analyses do not typically accounted for the anatomical distribution of values nor partial volume effects (PVE), contamination of binarised tissue masks with different vascular responses from other adjacent tissues [3]. In this project, we used anatomically-based simulations to assess the effect of partial volume contamination on CVR analyses.

We used an openly accessible dataset (https://doi.org/10.7488/ds/3492, n=15) to determine tissue-specific ground-truth CVR magnitude and delay distributions as briefly described below [2]. For each subject, we segmented the T1-w scan in to cerebrospinal fluid (CSF), cortical and subcortical grey (SGM) and white (WM) matter. We preprocessed the best quality BOLD scan and calculated voxelwise CVR using a delay-adjusted linear regression, as previously described [4]. As a proxy for blood vessels, which tend to have low BOLD signal and higher signal variability, [5] we calculated percentage voxelwise temporal noise-to-signal ratio and thresholded at 6%, based on visual inspection of the resulting masks. For each region, we defined CVR magnitude and delay distributions using probability density functions derived from the Gaussian kernel density estimate. We simulated CVR magnitude and delay maps by sampling from each tissue distribution using the high-resolution (0.5mm3) Multimodal Imaging-based Detailed Anatomical (MIDA) template as a reference to define the tissue type [6]. For each simulated map, we created a simulated BOLD timecourse (Fig.1) assuming a block-design end-tidal CO2 trace (2 minutes at 40 mmHg normocapnia, 3 minutes at 50 mmHg hypercapnia alternating for 12 minutes), replicating previous work [2]. For each voxel we shifted the timecourse by the simulated delay, added per-tissue signal intensity offsets based on the BOLD baseline signal and random Gaussian noise with temporal contrast-to-noise ratios matched to the relevant tissue. To introduce partial volume effect and reflect the spatial resolution of typical BOLD-CVR acquisitions, we downsampled the resulting simulated BOLD scans to the acquisition resolution (2.5mm3) using AntsPy with trilinear interpolation to apply a non-linear transform between the MIDA and subject-specific mean BOLD [7]. We calculated CVR using delay-adjusted linear regression [4]. For each dataset we performed 10 simulations. We visually compared the simulated maps against representative CVR magnitude maps from the reference dataset. Using Bland-Altman plots, we assessed per-tissue CVR magnitude differences between the simulated CVR map registered to the mean BOLD space and calculated CVR after downsampling the simulated BOLD timecourse. To assess the influence of partial volume effect, we plotted CVR magnitude and delay separately against the proportion of contamination from grey matter and vessels.

We found simulated CVR maps were comparable to the downsampled reference (Fig.2), After downsampling, CVR magnitude tended lower relative to the simulated reference map in all tissues, but Bland-Altman plots showed no systematic bias (Fig.3). As the proportion of blood vessel and grey matter partial volume contamination increased, simulated WM CVR magnitude increased and delay shortened, with a more pronounced effect for voxels adjacent to blood vessels (Fig.4A-B). For example, in voxels contaminated by surrounding blood vessels WM CVR magnitude was 59% higher than in uncontaminated voxels.

We showed simulations informed by underlying anatomy can be used to evaluate how partial volume contamination from surrounding tissues can affect CVR magnitude and delay. As well as identifying how partial volume effects may impact CVR analyses, such simulations can be used to evaluate robustness of different methods to contamination and pre/post-processing strategies to mitigate the impact [3, 8]. Future comparisons would benefit from more accurate blood vessel segmentations and should quantify how CVR is affected in voxels adjacent to CSF.

We successfully implemented anatomically-based computational simulations to assess the effect of partial volume effects on CVR magnitude, helping inform interpretation of CVR analyses and may help develop more robust CVR quantification methods.

Magnetic Resonance (MR) technique, commonly used for anatomical imaging, can also assess dielectric properties (DPs) of biological tissues. DPs describe tissues response to applied electromagnetic fields (EMF) and are essential for the planning, monitoring and optimization of medical techniques based on the use of EMF on patients [1]. MR-Electric Properties Tomography (EPT) approach reconstructs the DPs of tissues at the resonance frequency of the MR scanner from the amplitude and phase of the radiofrequency transmit magnetic induction B1 [2]. Due to the invasiveness of traditional techniques for the dielectric characterization of tissues, the possibility to use MR for this aim is very promising. MR-EPT workflow is made of two steps. First, the spatial distribution of B1 in the tissues is measured through specific sequences. Secondly, tissues DPs are reconstructed using different methods [3]. Both the sequence definition and the choice of the EPT method are complex tasks to manage, being dependent on the experimental case of interest. The aim of this work is to investigate MR-EPT sequences in the reconstruction of the conductivity of biological tissues. To this end, ex-vivo animal muscle and fat tissues were imaged, adopting three different sequences for phase mapping and exploring the differences among the three frameworks.

The MR acquisitions were done using a 3 T body coil scanner (Skyra, Siemens) for transmission and both the body coil and a head-neck 20 channels coil (Siemens) for reception. Muscle and fat ex-vivo samples were placed in a plastic beaker within the scanner. The B1 phase was measured with three sequences: a spin-echo (SE) based vendor-built sequence, called RF map, a classical spin echo, and a balanced steady state free precession (bSSFP) based vendor-built sequence, called BEAT. Raw data were acquired and processed to obtain the phase of the signal. The conductivity was derived at 128 MHz using the forward Helmholtz method implemented in MATLAB [3]. This method is straightforward to implement and physically based but, with respect to the other EPT methods, is strongly affected by noise and boundary errors. As reference for the MR-EPT obtained results, the conductivity of both tissues at 128 MHz was measured with the open-ended probe technique [4].

The phase map was derived from the raw data for each of the three sequences, considering reception both with the head and the body coil. The RF map sequence was characterized by an acquisition time of 8.53 minutes and by a slice thickness of 5 mm which could bring, in the in vivo case, to the loss of important anatomical details smaller than 5 mm. The SE sequence required 11.8 minutes for the acquisition but the slice thickness had a more acceptable value of 2 mm. The BEAT sequence lasted, instead, only 41 seconds and provided a slice thickness of 2.5 mm, but the obtained signal was affected by banding artifacts. These artifacts must be deleted through the tuning of the sequence parameters for the specific case of interest. Therefore, a proper shimming was applied, deleting these artifacts. Once the phase was obtained, the conductivity of muscle and fat was derived with H-EPT and some image processing procedures such as denoising and segmentation were applied. The modal value and the standard deviation among the values obtained in the pixels of the image were computed for each tissue and each sequence. The best agreement with the open-ended probe measurements was found with the SE sequence when the head coil was used as receiver, with percentage errors of 6 % and 2.50 % for muscle and fat, respectively. In the case of the RF map, errors of 13.4 % (muscle) and 0.83 % (fat) were obtained, while errors of 76.07 % (muscle) and 33.14 % (fat) were found for the bSSFP. For the body coil case, results obtained for fat were very distant from the open probe measurements probably due to the higher level of noise when this coil was used as receiver [5]. Overall, H-EPT results, considered with the associated variability, agreed with open probe measurements.

The results showed that, depending on the adopted sequence for phase mapping, some differences in the conductivity reconstruction are found. The use of a SE sequence adopting the head coil as receiver provided the best results but also the longest acquisition time. Even the RF map provided good results but long acquisition times and a too big slice thickness. On the contrary, the bSSFP sequence provided a faster acquisition with good slice thickness value but worse conductivity results and the need of tuning the parameters to delete banding artifacts.

This work compared three MR sequences applied using a Skyra Siemens scanner for the measurement of the B1 phase for the conductivity derivation through EPT. The findings demonstrated that, depending on the adopted sequence and coil for reception, specific operations must be considered adjusting the framework to find the best experimental compromise.

Diffusion-weighted steady-state free precession (DW-SSFP) [1] is a diffusion imaging sequence achieving high SNR-efficiency and strong diffusion-weighting [2] (Figure 1). Previous work has demonstrated that DW-SSFP signal measurements may be highly sensitive to microstructural features [3,4,5], motivating the use of sophisticated biophysical models for DW-SSFP investigations. However, the complicated signal-forming mechanisms of DW-SSFP [6] currently limits the integration of advanced simulations with parameter estimation routines [4]. In this work, I investigate the integration of DW-SSFP with Neural Posterior Estimation (NPE), a parameter inference technique leveraging concepts from Bayesian statistics and machine learning [7-9]. Briefly, given a prior distribution of parameters, P(θ), and a forward simulation model, f(θ→S), NPE uses simulated data pairs [D,S] to train a neural network to directly estimate the posterior distribution, P(θ | S). Once trained, experimental data can be passed to the network to perform parameter inference (Figure 2). Previous work has demonstrated the potential of NPE for diffusion-weighted spin-echo (DW-SE) investigations [10,11]. However, the DW-SSFP signal has additional dependencies on tissue relaxation properties (T1 & T2) and B1, which must be estimated for accurate diffusion modelling [6]. From the perspective of NPE, this corresponds to a different forward model per [T1,T2,B1] combination. Training a different forward model per [T1,T2,B1] combination is infeasible, requiring the adoption of an alternative method to address signal dependencies. Here, I address signal dependences in DW-SSFP by adapting an NPE network to estimate P(θ | S,T1,T2,B1), i.e. the posterior distribution conditioned on the measured signal and known [T1,T2,B1] values (Figure 2). The proposed network successfully accounts for DW-SSFP signal dependencies, achieving >300x acceleration and excellent agreement with conventional non-linear least-squares (NLLS). Evaluations are performed using a Tensor representation of the DW-SSFP signal.

DW-SSFP data was acquired in a single post-mortem brain (Siemens 7T; 0.85 mm iso.; 120 directions), alongside complementary turbo-inversion recovery (TIR), turbo-spin echo (TSE) and actual flip angle (AFI) [12] data for T1, T2 and B1 mapping. For full details of sequence parameters and mapping techniques, see [13]. A forward model of the DW-SSFP signal incorporating a Tensor [4] was established incorporating experimental DW-SSFP acquisition parameters, T1, T2, B1, and Tensor coefficients. NPE was implemented using the SBI toolbox (0.23.3) [14] in Python (3.12.8) as follows: - Priors: Uniform distributions, limits = [0, 0.5] μm/ms (diagonal components) & [-0.25, 0.25] μm/ms (off-diagonal components). A classifier network was trained [15] to ensure simulated Tensors were positive semi-definite. - Data Pairs: 6,000,000 data pairs, corresponding to 1,000,000 simulations with 5 Rician noise levels (uniform distribution, SNR limits = [2, 50]) + noise free. Each simulation was associated with an arbitrary Tensor (prior estimate), T1, T2 and B1 (uniform distributions, limits [300, 1200] ms, [20, 80] ms, and [0.2, 1.2]). - Training: The NPE network was implemented using neural spline flows [16], with default SBI toolbox parameters. [T1,T2,B1] conditioning was achieved by appending their values to the signal during training, creating a 1D vector corresponding to [S,T1,T2,B1] (Figure 2). The network was trained on a personal laptop in ~31 hours, converging after 210 epochs. - Evaluation: The network was evaluated using both simulated and experimental DW-SSFP data, performing comparisons with NLLS.

Figure 3a-c compares estimated Tensor coefficients as a function (a) T1, (b) T2, and (c) B1 using NPE. Excellent agreement is found with ground truth (0.2% mean difference). Figure 3d compares the accuracy of parameter estimation for NPE and NLLS associated with different SNR levels, achieving similar accuracy (0.3% mean difference). NPE benefits from ~380x acceleration versus NLLS for parameter estimation. Figure 4 compares experimental Tensor estimates using (a) NPE and (b) NLLS. Excellent agreement is found, achieving R = 0.974 for fractional anisotropy (FA) maps. NPE additionally benefits from detailed posterior distributions, with Figure 4c displaying an example distribution for a single corpus callosum voxel.

I demonstrate that NPE achieves fast and accurate parameter inference from DW-SSFP data, with the proposed conditioning approach accounting for dependencies on T1, T2 and B1. Resulting parameter estimates give excellent agreement to ground-truth simulations (Figure 3) and experimental NLLS estimates (Figure 4). The implemented routine could be adapted to incorporate more sophisticated microstructural models, including Monte Carlo simulations [17].

NPE achieves fast and accurate parameter inference for DW-SSFP investigations.

Diagnostic tests are key medical tools to detect or rule out specific pathologies. Magnetic Resonance Imaging (MRI) is very attractive due to its non-invasive and non-ionizing nature, but its high cost, complex infrastructure and long scanning times, keep it enclosed in hospitals and tied to long bottlenecks [1]. Low Field (LF) MRI relaxes cost and hardware specifications at the expense of image resolution, often improved by AI post-processing [2], [3]. Yet, insofar as MR is image-based, hardware and scanning time necessities remain hefty. Whereas previous imageless attempts detaching image from MR still worked on k-space data representation [4]–[6], in this work we advocate for a more radical Imageless MR Diagnosis (IMRD), based on raw 1D MR time signals, further lowering barriers for MR use. An IMRD solution comprises (i) the minimal structure of hardware, (ii) the optimal information encoding in MR data and (iii) optimal models for processing such data. In this work, we showcase an IMRD proof-of-concept proposing a combination of these three components with a simulated case study, aiming to detect and quantify in silico Multiple Sclerosis (MS) lesions [7], [8].

We downloaded 17 healthy phantoms with White Matter (WM), Grey Matter (GM) and Cerbrospinal Fluid (CSF) tissues from Brainweb [9], and in-silico MS lesions were simulated in each phantom [10]. Healthy and MS-affected slices were randomly selected to build a final data set of 935 brain slices, with in-silico MS in nearly 40 % of them (Figure 1). Next, we simulated MR signals for each slice in the data set (Figure 2). The simulated acquisitions used a ZTE-like sequence [11]. Once a radial spoke is set, an initial Inversion Recovery (IR) pulse is followed by a train of RF pulses, introducing a rewind gradient at half the TR. The time for IR (∈[0.01,2] s), the Repetition Times (TRs, ∈[10,500] ms) and Flip Angles (FA ∈[10,150] º), of the MR sequence were optimized to maximize MS tissue distinguishability by minimizing the cross-correlation between tissues’ transverse magnetization signals. Literature T1 and T2 values for WM, GM, CSF and MS at 1.5 T were used [12]–[14]. Besides, we considered two minimal-hardware acquisitions: single-gradient (Figure 2 b)— with 1-dimensional spatial encoding by applying a frequency gradient in the X-axis at 0º—and gradientless (Figure 2 c)—with the pure Free Induction Decay (FID) signal. The final sequence consisted of 30 TRs and takes less than five seconds in both acquisitions. White noise at 54 dB was injected into the simulated MR signals, a reference value extracted from our portable 72 mT scanner [15], [16]. Finally, we trained 1D Convolutional Neural Networks (1D CNNs, Figure 3) [17] to estimate the volume and presence of MS lesions from each acquisition’s 1D MR signals, leaving a test set of 110 slices from two phantoms for final model evaluation.

Figure 4 shows the results in the test set. The single-spoke acquisition reported an MS lesion detection AUC of 0.95, and a volume estimation accuracy of R² close to 0.8. The gradientless setup yielded an AUC of 0.8 and an R² of 0.98. The largest missed lesions were 0.06 mL and 0.26 mL for single-spoke and gradientless, respectively.

The more accurate MS volume estimation with gradientless suggests that the dephasing induced by the gradient in single-spoke accelerates the decay of transversal magnetization signals (Figure 2 b), governed by T2*, losing the relevant information that is otherwise preserved in the gradientless’ signal slower decay (Figure 2 c). Nonetheless, the overestimation at zero MS volumes, (Figure 4 d), still leaves room for improvement, and real-world MS cases are likely to require additional data processing steps.

IMRD relies on raw MR signals optimally encoding information registered by minimal hardware and scan time. Our in-silico case study aimed to exemplify such an IMRD framework in practice by detecting or quantifying MS tissue within a sample. Results with this in silico prototype of IMRD solution with (i) minimal or no spatial encoding, (ii) fast MR sequences below 5 seconds and (iii) data-driven models as CNNs, suggest that IMRD frameworks could successfully answer closed diagnostic questions. Yet, different scenarios—detecting other pathological events, as stroke or liver fibrosis, or measuring the spatial features of tissues as in lissencephaly—will probably require a different articulation of components. It is therefore vital to test IMRD with more clinical enquiries, ex vivo and in vivo data, which will probably imply changes in hardware, MR sequence and data processing. If successful, IMRD approaches could widen the possibilities of LF MR scanners, bringing on-site and nearly instant MR-based diagnosis.

Diffusion MRI (dMRI) provides valuable insight into tissue microstructure for clinical and research applications. Traditional diffusion tensor imaging (DTI) models diffusion as a Gaussian process, which limits its accuracy in regions with complex fibre configurations (such as crossing or bifurcating tracts) [1]. Higher-order models, including diffusion kurtosis imaging (DKI) and constrained spherical deconvolution (CSD), address these limitations by capturing non-Gaussian diffusion behaviour or by resolving multiple fibre orientations within a voxel. Despite their theoretical advantages, these models are more sensitive to noise and acquisition variability, raising concerns about reproducibility [2]. Currently, there is no standardized method to perform quality assurance on dMRI data, and there are limited studies measuring the reproducibility of higher-order tensor metrics. This study aims to investigate the consistency of a novel anisotropic diffusion phantom (PreOperative Performance, Toronto, ON) for higher-order tensor dMRI protocols. While this phantom has recently demonstrated reliability for rank-2 tensor metrics across different MRI vendors, its suitability for higher-order models remains to be established [3].

The phantom consists of fixed cylindrical synthetic filaments (diameter = 2 μm) embedded in a PVA matrix, arranged into modules that emulate linear, crossing, and branching white matter structures. It was scanned 11 times across 3 days on a 3T Discovery MR750 (General Electric HealthCare, Waukesha, WI). The protocol included a conventional DTI sequence (30 directions, b=1000 s/mm²), two high angular resolution diffusion imaging (HARDI) acquisitions (60 and 90 directions, b=1300 s/mm²), and a multi-shell DKI sequence (30 directions, b=250–3000 s/mm²). Six regions of interest (ROIs) were drawn in various locations (Figure 1). Metrics were extracted using DIPY and FSL from three models: DTI (FA, MD, AD, RD), DKI (KFA, MK, AK, RK), and CSD (GFA). Reproducibility was assessed using coefficient of variation (CoV) and intraclass correlation coefficient (ICC).

DTI-derived metrics showed excellent reproducibility across all ROIs (ICC > 0.9), with HARDI acquisitions slightly outperforming DTI in reducing variability (Figure 2). FA exhibited CoVs < 10% while MD, AD, and RD had CoVs < 3%. DKI metrics displayed greater variability (Figure 3). KFA showed moderate CoV (~4–11%) but maintained strong ICCs. MK values ranged widely in CoV (10–27%), with the highest variability in ROIs featuring fibre crossings or branching. AK and RK had the least stability, with CoV exceeding 30% in certain ROIs, reflecting model sensitivity to fitting instability at high b-values. GFA, derived from CSD, had moderate reproducibility across acquisitions (Figure 3). CoV was lowest for HARDI-60 (2.30%) and highest for DTI (3.26%). ICCs were 0.6603 (DTI), 0.8046 (HARDI-60), and 0.8507 (HARDI-90), suggesting increasing reliability with higher angular resolution (Figure 3).

The phantom showed high reproducibility for DTI metrics, supporting its use in conventional tensor-based QA. HARDI protocols provided modest improvements in complex regions, aligning with research showing increased angular resolution enhances tensor fitting [2]. In contrast, DKI parameters demonstrated moderate reproducibility, with KFA being the most stable (CoV ~7.36%, ICC 0.9361). MK, AK, and RK showed increased variability (CoV 14-18%), particularly in complex fibre regions, reflecting the sensitivity of kurtosis metrics to noise and acquisition parameters. These findings suggest DKI metrics require more stringent quality assurance protocols when used for quantitative assessment. CSD-derived GFA exhibited improved reproducibility with increased angular resolution, as evidenced by the higher ICC values for HARDI-90 (0.8507) compared to conventional DTI (0.6603). Importantly, qualitative examination of CSD fibre orientation distribution functions (Figure 4) demonstrates the phantom's ability to accurately represent complex fibre architectures that cannot be resolved using conventional DTI. Distinct peaks were observed at expected crossing locations, suggesting the phantom’s internal geometry provides sufficient angular contrast to evaluate orientation-resolving models. However, the variability in GFA and DKI metrics highlight the need for further refinement of acquisition and processing pipelines before routine use of higher-order models in QA workflows.

The PreOperative Performance phantom offers a consistent and repeatable platform for assessing conventional dMRI metrics and shows promise for evaluating more advanced modelling frameworks. These findings support the use of this phantom in future multi-centre studies but also emphasize the continued need to refine acquisition protocols and modelling techniques for higher-order diffusion analyses. Future work should expand to cover inter-scanner and inter-vendor reproducibility and evaluate phantom stability over time.

Quantitative Susceptibility Mapping (QSM) is a quantitative MRI technique that maps the spatial distribution of magnetic susceptibility, having shown promise for biomarker imaging applications [1]. For QSM to reach widespread clinical use, its acquisition time must be reduced whilst maintaining accuracy and image quality. Parallel imaging (PI) techniques are extensively used to reduce clinical acquisition time. However PI must be carefully optimised as higher levels of acceleration can degrade image quality by introducing heterogenous noise and aliasing artefact [2–4]. Phantoms can provide reference values for evaluating the performance of quantitative MRI techniques [5, 6]. This study aims to assess the impact of PI acceleration, for reducing QSM acquisition times, on the accuracy and quality of QSM maps.

A Gadolinium-based phantom was constructed to provide reference susceptibility values. The phantom consisted of a 2L water-filled container with six targets containers (diameter=25mm, length=92mm), five with varying concentrations of a Gadolinium-based contrast agent (Clariscan 279.3 mg/ml, gadoteric acid, GE Healthcare), mimicking susceptibility values in-vivo (0.052, 0.104, 0.155,0.31,0.57ppm) [7] and one containing water only. Multi-echo gradient echo magnitude and phase images were acquired at 1.5 T using recommended acquisition parameters from the QSM consensus group (see figure 1B) [8]. Parallel imaging factors were varied both in-plane (R1) and through-plane (R2), with a total of five acquisitions obtained. The acquisition with PI factors R1=2 and R2=1 was taken as the “gold-standard”. QSM maps were reconstructed using the SEPIA Toolbox in MATLAB (2024b) [9, 10]. A “two-pass masking” based approach was performed [11] using nonlinear echo-combination[12], phase unwrapping using SEGUE[13] , background field removal using PDF[14] and dipole inversion using Iterative Tikhonov [15]. Six regions of interest (ROI) were defined in the phantom targets[16]. Mean and standard deviation susceptibility were calculated for each ROI and acquisition and plotted against expected susceptibility. Agreement was assessed using linear regression to obtain the slope, intercept and coefficient of determination (R2). The modified structural similarity index measure (SSIM) for QSM, XSIM, was used to compare image quality against the gold-standard acquisition, yielding four XSIM maps [17]. Parameters for calculating the XSIM maps (K1 = 0.01, K2 = 0.001 and L=1) were obtained from Milovic et al., 2022 [18].

PI accelerated QSM maps are shown in figure 2, illustrating the qualitative differences between acquisitions. Measured versus expected susceptibility values are shown in figure 3. Vertical error bars represent ROI standard deviations while horizontal error bars represent Gadolinium concentration uncertainty of each target. Solid lines (blue) show estimates from linear regression with dotted lines showing ideal agreement. Shaded areas are bounded by 95% confidence intervals of the fit. Fit parameters are included with associated 95% confidence intervals. In-plane, through-plane XSIM maps and mean XSIM for PI factors are shown in figure 4.

Quantitative susceptibility measurements (figure 3) show good linear agreement with expected values for QSM maps with increasing PI factor: (R1=2, R2=2), (R1=3, R2=1) and (R1=3, R2=2). The results show the potential to reduce acquisition times to 33% that of standard scanning while maintaining accuracy. However, quantitative measurements are impacted when using parallel imaging factors of R1=3 and R2=3, as severe artifacts are apparent. XSIM measurements demonstrate a decrease in image quality with increasing parallel imaging factor. XSIM maps are heterogeneous, highlighting the non-uniform increase in noise for increasing parallel imaging factor.

This study assessed the effect of increased parallel imaging on measured susceptibility values and image quality. Good agreement was found between ROI measurements and expected values despite increasing in-plane and through-plane PI factors. However, a parallel imaging factor of 3 for both R1 and R2 yielded severely artifacted QSM maps and poor susceptibility measurements. XSIM indicates decreasing image quality with increasing PI acceleration and highlights the heterogeneity of noise introduced by PI. Further work will look at the impact of partial fourier imaging on the accuracy and image quality and develop diamagnetic targets to characterise the full spectrum of susceptibility found in vivo.

While quantitative water-specific T1, R2* and fat-fraction (FF) mapping is of great interest in liver imaging [1,2], conventional methods are typically time-intensive, since they require individual data acquisition for each map. Here, we apply a fully non-linear model to reconstruct water-specific T1, R2*, and B0 field maps directly from k-space data [3]. By evaluating results obtained from our in-house designed and manufactured water-fat-SPIO phantom, we present one step towards validating our proposed method. Here, the goal was to design and construct a quantitative MR phantom that covers all three parameter maps with good accuracy while keeping the manufacturing process simple.

When designing our water-fat-SPIO phantom, we adopted manufacturing instructions of existing protocols [4,5]. Our phantom is comprised of 18 vials (18 mL) with varying designed fat volume percentages (0, 5, 10, 20, 40 and 100%), iron concentrations (0, 50, and 150 µg/mL) and T1 water values (800 and 1500 ms). Similar to Hines et al. we chose peanut oil and super paramagnetic iron oxide (SPIO) particles (magnetite, 5 nm, Sigma-Aldrich) to modulate fat concentration and R2*. Distilled water was doped with gadobutrol (Gadovist, Bayer Vital GmbH, Germany) to modulate water-specific T1 values. For gelling, we added Agar (2.23% w/v) over heat while stirring. All vials were placed in a 1 liter cylindrical container in two sets (L1, L2). Background was filled with distilled water. To achieve fast multi-parameter acquisition, we combine a single-shot inversion recovery (IR) sequence with a continuous multi-echo radial FLASH readout and incorporate blip gradients between echoes for improved k-space coverage [6] (Fig. 1). To jointly estimate water-specific T1, R2*, B0, and FF maps directly from the acquired k-space data, we model the underlying physical signal and formulate parameter estimation as a non-linear inverse problem. The model explicitly accounts for water and fat specific equilibrium and steady-state signal contributions and for their effective longitudinal relaxation rates. In addition, we consider the 6-peak fat spectrum [7] and field inhomogeneity. The forward model accounts for the radial sampling pattern, coil sensitivities and Fourier transform. The optimization problem is solved iteratively using IRGNM-FISTA [8-11] implemented and performed in BART [12], utilizing Sobolev and ℓ1-wavelet regularization. Water-fat-SPIO data was acquired on a 3T Siemens Magnetom Vida equipped with a 20 channel head coil.

First, we validated the proposed model-based reconstruction on a numerical phantom (BART) (Fig. 2a), covering a wide range of ground-truth values. Comparison of ROI-averaged simulated values to the ground truth shows overall low mean differences (Fig. 2b). In physical phantom studies, reference maps for R2*, FF, and B0 were estimated using model-based reconstruction of steady-state multi-echo data (0.8x0.8x5 mm3) [13]. Steady-state data was extracted from the last 140 excitations of the same acquisition. T1 references were estimated using single-shot IR FLASH with single-echo readout [11]. Figure 3 shows reconstructed maps from one set of vials. To assess accuracy, ROI averaged mean values of reference and the here proposed method were compared. Resulting maps of water-specific T1, R2* and FF are in excellent agreement, as indicated by Pearson correlation coefficient. Differences in B0 maps need further investigation.

To estimate manufacturing accuracy, we compared designed ground truth values to values obtained from our proposed method (Fig. 4, top). Although designed fat-fraction values were achieved adequately, we noticed deviations from the designed values in both R2* and T1. While deviations in R2* can be attributed to varying fat-fractions and its know effect on measured R2* values [5], discrepancy in T1 values can only be explained by manufacturing inaccuracies at this point. To rule out acquisition errors, we additionally performed gold-standard T1 mapping [14] using IR spin-echo water-only excitation acquisitions with varying TIs (TI = 30, 280, 530, 780, 1030, 1280, 1530 ms). We can report excellent agreement between gold-standard T1 maps and our proposed method (see Fig.4, bottom).

We designed and manufactured a simple water-fat-SPIO phantom used towards validating our new proposed method. In our proposed method, we combined a non-linear model-based reconstruction with radial IR multi-echo FLASH acquisition enabling joint estimation of water-specific T1, R2*, FF, and B0 field maps from a single-shot acquisition of four seconds while maintaining high accuracy. Our developed method can potentially improve patient comfort and add valuable diagnostic information while simultaneously rendering multi-parametric quantitative MRI more feasible for clinical applications.

In MRI scans using a closed-bore MRI system, patients are put into the bore either head-first or feet-first, mainly based on the anatomical region being imaged, ensuring that the target area is centered and that patient comfort is maximized. Beyond these practical considerations, the interplay between the B0 field orientation (or equivalently, patient orientation) and the RF system may introduce variations in the imaging process. In this work, we investigated the influence of patient orientation on ultimate intrinsic SNR (uiSNR)[1], which is an intrinsic feature of RF reception unaffected by receive array design.

Realistic body model: The Duke body model from the Virtual Family (IT’IS Foundation)[2] was used for electromagnetic simulations. Frequency-dependent properties[3] were assigned to different tissues in the body model. B0 strengths of 1.5 T, 3 T, 7 T, 10.5 T, and 14 T were included in the simulations. Additionally, a homogeneous body model with average tissue properties and a mirror-symmetric body model were constructed from the Duke model and assigned tissue properties at 7 T. All body models were truncated from the neck to the thighs to reduce the computational burden (Figure 1). Basis set construction: For each simulation, an electromagnetic basis set containing 3000 random bases was generated following the method of Guerin et al.[4]. Each basis contains an electric field and its corresponding magnetic field. uiSNR calculation: The uiSNR was calculated in each voxel using the electromagnetic basis set [5], which involved computing B1-, the negatively rotating component of the magnetic field in the plane perpendicular to B0. The uiSNR maps were computed for two opposite B0 orientations (head-first and feet-first) and compared by calculating ratio maps.

Figure 2 shows the uiSNR ratio maps between the two opposite B0 orientations in the original, the homogeneous, and the mirror-symmetric Duke model at 7 T. In the homogeneous model, regions where the uiSNR increases or decreases when B0 is reversed can be observed. These regions are relatively large and mainly appear in areas where adjacent body parts are separated by air. In the original heterogeneous model, regions with increased or decreased uiSNR appear on a smaller geometrical scale. In the mirror-symmetric model, regions with increased and decreased uiSNR always appear symmetrically in pairs. In the original model, similar patterns can also be observed to some extent in regions that are roughly symmetric. Figure 3 shows the uiSNR ratio maps of the Duke model at different B0 field strengths. From 1.5 T to 14 T, the extent of variation between the two B0 orientations increases, which can also be observed from the histogram of uiSNR ratio (Figure 4). At 14 T, the difference in uiSNR caused by the two opposite B0 orientations can be as large as 50%.

Model limitations: With the random basis set used in this work, convergence of SNR towards uiSNR was only achieved in regions deeper than about 3 cm from the body surface. Effect of B0 orientation: Differences between uiSNR maps obtained with two opposite B0 orientations are a combined result of overall shape and heterogeneous local structures. For some positions in the body, there may exist a preferred orientation of B0. As ultra-high-field MRI exploits higher field strengths, this effect may become more relevant in the future as a limiting or boosting factor of image quality. Transmission vs. reception: Transmission and reception are intrinsically linked, representing complementary aspects of the same physical process. If a position has a preferred orientation for signal reception due to global and local geometries, the B0 orientation may offer an advantage or disadvantage for excitation.

In this work, we illustrated an interesting effect that patient orientation (or equivalently, B0 orientation) may influence the uiSNR in a realistic human body model. The extent of this effect increases with B0 strength. B0 orientation is therefore proposed as an imaging factor to be considered as B0 strength increases into the ultra-high-field range.

Quality assurance (QA) is essential to ensure consistent performance of MRI scanners, especially in multicenter studies where inter-scanner variability can impact data comparability [1,2]. Using phantom-based acquisitions allows for quantitative MRI (qMRI) assessment without biological confounds, enabling effective monitoring of scanner stability and temporal drifts [3,4]. To support QA harmonization, the Italian Ministry of Health established the RIN – Neuroimaging Network [5], comprising 23 Scientific Institutes of Hospitalization and Care (IRCCSs) working on unified protocols for acquiring, processing, and sharing qMRI data. Functional MRI (fMRI), being highly sensitive to signal fluctuations, particularly benefits from reliable QA [6,2]. In this context, we present a harmonized, multivendor QA toolbox aimed at assessing intra-site stability and inter-site reproducibility through phantom data, producing automated visual and numerical reports to support reproducibility across centers.

A QA study was conducted across 16 Italian RIN centers equipped with 3T scanners from three vendors. From 2017 to 2022, centers performed monthly scans using the FUNSTAR phantom [7] and a standardized fMRI protocol. The open-source Python-based RYAN toolbox [8,9] was used to analyze key QA metrics—SNR, SFNR, signal drift, Weisskoff radius of decorrelation (RDC), and even-odd variance—across central and peripheral ROIs. Metric trends were monitored over time and statistically compared across vendors and centers. Acceptance thresholds were defined for each metric and visualized using Bland-Altman plots [10]. Spike detection was performed [11], and intraclass correlation coefficients (ICCs) were computed to assess scanner reliability [12,13].

Metrics were visualized via boxplots, barplots, and coefficient of variation (CV) analyses. SNR and SNRt showed no central vs peripheral ROI differences, but significant inter-center variability [14,15], with outliers affecting SNR results (Fig.1,Fig.2a,Fig.2b,Fig.3a,Fig.3b). SFNR had high ICCs centrally (>0.9) but lower peripherally (Fig.2c), with vendor-dependent effects [16] (Fig.1a). Signal drift varied across ROIs and vendors, with moderate ICCs (0.3–0.4), especially in peripheral regions [17,18] (Fig.1,Fig.3c). Weisskoff RDCs correlated well across 2D and 3D planes [19] (Fig.1c,Fig.4a, Fig.4b, Fig.4c); the 3D method proved useful for detecting inter-slice correlated noise [20] (Fig.2e). Even-odd variance had low variability (ICC 0.7–0.8) (Fig.2f,Fig.3f)), though vendor-related differences were noted(Fig.1a). Over five years of longitudinal tracking, SNR, SNRt, and signal drift exhibited greater variability, whereas SFNR, RDCs, and even-odd variance remained comparatively stable [21] (Fig.4d). Spike and Weisskoff plots helped assess signal stability, and most centers remained within QA thresholds with minimal exceptions.

This study introduces the RIN-Neuroimaging Network’s approach to standardize fMRI data acquisition across different vendors and sites using a common QA protocol and phantom. Key metrics—SNR, SFNR, drift, RDC, and even-odd variance—were used to evaluate performance variability. Peripheral ROIs showed more instability, especially for drift, likely due to gradient heating effects [22,23]. The Weisskoff RDC 3D extension helped identify slice-related noise [20]. Even-odd variance correlated inversely with RDC, supporting its role in quick noise estimation. ICC values confirmed high reproducibility, particularly for SFNR central and even-odd metrics [13]. Most centers operated within set thresholds, supporting scanner stability and effective QA implementation.

We developed the RYAN toolbox as part of a three-year QA program in the RIN-Neuroimaging Network [5]. Monthly assessments of key metrics (SNR, SFNR, drift, RDC, even-odd) provided insights into scanner performance. RYAN runs automatically on the shared FUNSTAR database and delivers detailed reports for each center, allowing vendors to be notified in case of deviations. It visualizes signal time-courses, spikes, and threshold violations, using acceptance ranges derived from 16 scanners across three vendors. Regular QA tracking helps identify hardware/software-related changes affecting scanner performance. Future developments will include enhanced SNR estimation using 32 background ROIs and a user-friendly GUI for on-site QA checks.

Tissue microstructure plays a central role in both clinical and neuroscience research. Currently, Diffusion Tensor Imaging (DTI) is the gold standard for the non-invasive assessment of brain white matter architecture. Despite so, conventional DTI metrics may not fully exploit the textural information embedded in diffusion-weighted images. Radiomics, a quantitative imaging approach, offers a powerful framework for extracting high-dimensional features that characterize image heterogeneity and complexity [1]. As a proof of concept, in this work we demonstrate that radiomic features can describe diffusion properties using a reference phantom specifically designed to simulate microscopic diffusion anisotropy on clinical MRI systems [2, 3].

The analysed phantom presented two concentric NMR-tubes, with inner and outer diameters of 5 mm and 10 mm, respectively. The central tube was filled with a reverse hexagonal-type liquid crystal, consisting of nanometre-scale water channels embedded within a continuous matrix of detergent and hydrocarbon, while the outer one was filled with a polymer solution. To randomize the orientation of the liquid crystal’s domains the sample was initially melted and, subsequently, domains alignment along a single direction was induced by exposing the phantom to an 11.7 T magnetic field for seven days. A total of 110 phantom acquisitions were obtained at evenly spaced time intervals from which multiple diffusion maps were computed, including Fractional Anisotropy (FA), Micro-Fractional Anisotropy (µFA), Mean Diffusivity (MD_xx, MD_yy, MD_zz), mean Isotropic Diffusivity (D_iso), and mean squared normalized Anisotropy (D_Δ^2) [4]. Five regions of interest (ROIs) were segmented using a threshold-supervised algorithm: the pure liquid crystal region (LQ), the pure polymer region (POL), and an intermediate zone (Mix) that includes contributions from both. In addition, two composite regions were defined: LQmix, obtained by merging LQ and Mix; and POLmix, obtained by merging Mix and POL. After pre-processing, 98 radiomic features were extracted from each diffusion metric–ROI combination using PyRadiomics [5]. Features statistically associated with liquid crystal domains alignment were identified using Spearman’s rank correlation coefficient (ρ), with significance defined as p-value < 0.01 after Bonferroni correction for multiple comparisons.

A total of 3430 feature-metric-ROI combinations were analysed. Among these, 1723 resulted to be significantly related with liquid crystal orientation. Of these, 867 features showed a strong correlation (|ρ| ≥ 0.7), while 856 demonstrated a moderate correlation (0.3 < |ρ| < 0.7). Strongly related features were predominantly derived from the LQmix region (279/867), Mix region (222/867) and LQ (202/867) pure region. In contrast, the 60% of features not related with the alignment were associated with POL and POLmix regions. Considering the MRI metrics, strongly related features were from the majority obtained from FA (24%), D_Δ^2 (20%), MD_zz (20%), followed by MD_xx (11%), MD_yy (10%), D_iso (9%), and µFA (6%).

Among the pure regions, LQ was expected to exhibit the highest significant correlation with domains alignment, as it was experimentally designed to display anisotropic behaviour. Conversely, the POL region showed isotropic diffusion across all time-spaced acquisitions, as reflected by the large number of features unrelated to domains' alignment. Focusing on MRI metrics, FA and D_Δ^2 were most frequently associated with alignment. FA quantifies the degree of diffusion anisotropy at the voxel level, while D_Δ^2 describes the anisotropic component of the diffusion tensor. Focusing on MD, the alignment effect was more evident in MD_zz than in MD_yy, MD_xx, reflecting sample geometry. Lastly, µFA showed the lowest number of strongly related features as it characterizes the anisotropy of the underlying microscopic structure but not it’s organization on the voxel level. A visual comparison of the grey-level distributions between the first and final acquisitions clearly reveals a substantial variation in the FA maps: mean FA values increased from 0.11 ± 0.02—indicative of near-isotropic diffusion—to 0.87 ± 0.03, reflecting a marked enhancement in anisotropy caused by the domain’s alignment. In contrast, the µFA maps exhibited minor variation, with the mean value remaining nearly constant (from 0.97 ± 0.01 to 0.99 ± 0.00), in line with theoretical expectations.

In this work, we aimed to assess the feasibility of radiomic analysis of diffusion MRI metrics using a well-characterized phantom. The results of the statistical analysis of radiomic features were largely consistent with the expected physical behaviour of the system, demonstrating that radiomics can effectively capture changes in the context of diffusion MRI.

Subject-specific quantitative maps can be obtained using conventional quantitative MRI or MR fingerprinting. Our approach uses a fast, low-resolution quantification scan in combination with an artificial neural network. Synthetic MRI contrasts are simulated using the derived quantitative maps (PD, T1, T2, T2’, D, dB0, B1) using an approach known as synthetic imaging11,12. We used the subject-specific synthetic images for the purpose of contrast previews and slice planning. In this abstract we put such synthetic MRI to the test by comparing against real scans of the same sequences using suitable metrics.

The quantification sequence was based on the prepared snapshot FLASH as proposed by Weinmüller1 et al., which we extended to whole brain coverage (2x2x4 mm3, GRAPPA6 - 6) and acquisition-time of 8 mins, while acceleration to 4 mins is possible - data not shown. The series of acquired images with PD-, T1-, T2-, T2’-, D-, dB0-, and B1+preparation is then quantified using a 3D convolution neural network trained with simulated images for same preparations using brainweb2 phantoms All sequences were programmed in the Pulseq standard9, which then directly used in the scanner as well as the simulation made the comparisons possible. Three typical clinical protocols chosen were the following: 1. Steady-state FLASH3 (as used in Localizer scans or SWI, with TE=1.37ms, TR=3.15ms, FA=8 deg, 1.6x1.6x1.6 mm resolution, 2. MPRAGE4 with TE=3.2ms, TR=8.16ms, FA=9 deg, TI=9ms, resolution=2 x2 x4 mm, and 3. T2w-TSE5 TI=1.11ms,TR=2.53ms, FA=90 deg ,resolution 2x2x4 mm. These protocols along with quantitative scan were performed on the volunteers at a 3T MAGNETOM Cima.X Scanner (Siemens Healthineers AG, Forchheim, Germany) after well informed written consent. Quantification maps of a volunteer was obtained with Inference time of 2s in seconds in Nvidia-RTX A4000 16 Gb GPU. For fast synthetic previews we used the Phase Distribution Graph (PDG) simulation7. Simulation time of these sequence on GPU for 3D Steady-state FLASH: 732s, 2D-MPRAGE: 128 s,2D- T2w-TSE:62s, both simulated and measured data was processed with the same reconstruction pipeline. Images were windowed to reflect similar GM/WM/CSF contrast. Since the experiment is validated with synthetic and real data of the same subject, we used structural similarity10 of equivalent 2D slices of interpolated resolution without any form of co-registration. MRI vendors offer automatic slice planning algorithms based on a 3D localizer overview scan of the scanned anatomy. In a second experiment, we evaluate the performance of the vendor’s automatic slice planning algorithm on simulated localizer images. A confidence score derived by the algorithm indicates the quality of the predicted slice planning.

Figure 1-3 show the visual comparison of our synthetic images based on the 2x2x4 mm quantification, together with the real acquisition of the same sequences, i.e. steady-state FLASH (Fig 1), MPRAGE (Fig 2), T2w-TSE (Fig 3). The volumetric confidence for Brain and Orbital part of 3D volume has least uncertainty than the real measurement as shown in the Table 1.

Due to trade of in acquisition time our previews suffer from low resolution and anisotropic resolution, which leads to partial volume blurring effects prevalent in both sagittal and transversal views. Apart from that, the contrast features of all sequences were replicated evidently. large deviations are visible in the fat signals of the FLASH steady state, which we traced back to T1 mismatch in the quantification. Image metrics (at the bottom of the Fig(1-3)), and volumetric confidence values reveals that synthetic images are close to real measurement, bringing us a step closer to potential preview and patient-specific process planning.

We showed that within the existing framework of MR-zero8, patient-specific synthetic preview is possible with reasonable (scan, inference and simulation) time and realistic contrast. Resolution and scan time of quantification can still be optimized with further accelerations. While similar approaches based on MR fingerprinting exist11,12, this work validates the generated quantitative digital twins for fast synthetic preview of digital twin of clinical sequences. In principle, synthetic preview can also be used with any other quantification method, if the complete list of parameters are provided.

Radiomics offers a non-invasive tool for assessing changes in digital images by extracting quantitative features, which can be used with other available information to assist decision-making(1). Radiomics has been used in studies such as survival and recurrence estimation for many tumours(2,3). The phantoms have been used to assess the MRI imaging chain's performance and thus have found applicability in assessing radiomic feature reproducibility and stability due to the ease of multiple imaging. The stability of radiomics features has been shown to depend on various factors, including changes within the material under study. Heterogeneous phantoms comprising various tissue-equivalent materials to provide realistic anatomy and internal tissue structures have become attractive in assessing the stability of radiomic features under various conditions. Heterogeneous phantoms comprising various tissue-equivalent materials to provide realistic anatomy and internal tissue structures have become attractive in assessing the stability of radiomic features under various conditions. Several phantoms have been proposed in the literature to assess the stability of the radiomic features, including fruits, vegetables, and 3d printed objects, among others(4–6). The selection of a phantom to use depends on the study's objective; for example, longitudinal studies require a phantom that does not change or deteriorate with time, as this will affect the radiomics feature(5). In this study, a polymer gel phantom was chosen as a possible candidate for use in MRI radiomic feature analysis. The gel phantom has the advantage of other objects being inserted inside it without losing its integrity(7). The polymer gel phantom's stability must be determined before it can be considered. The study aims to determine the stability over time of polymer gel phantom using radiomic texture features.

A gel phantom was constructed by using the (7) approach by using Carbopol-974p powder combined with distilled water, Mn(NO3)2, and NaOH, and the mixture poured into a test tube. A jig was made to have five test tubes be suspended in air as shown in figure 1. The phantom was imaged (3.0T Phillips Ingenia) bi-weekly over a period of 12 months and pulse sequence, T2w TSE was used and the acquisition parameter is shown in Table 1. A radiomic pipeline suggested by IBIS was used for feature extraction, and the PyRadiomics software of 3D Slicer was used for segmentation and feature extraction. Pre-processing normalisation was performed using a z-score on all images, and features were determined for the segmented volumes for each acquisition. For each MRI image, a total of 104 features were extracted: shape features, first-order features and second-order features. The second features are further divided into: gray level co-occurrence matrix (GLCM), gray level dependence matrix (GLDM), gray level size zone matrix (GLSZM), gray level run length matrix (GLRLM), and neighbouring gray tone difference matrix (NGTDM). Coefficient of Variance (CV) was used to determine texture features useful for further analysis and the threshold was <10%.

Preliminary results indicated that the 5 texture features of GLSZM showed repeatability over time as shown in table 2. While most of the shape features provided had the best CV values i.e <10%, they provide no information about the stability of the gel phantom. The GLCM and first-order each had one feature that was repeatable on all test tubes. The CV values for GLRLM and NGTDM features for all the test tubes were greater than 10% indicating they might no be useful in this case.

Preliminary results indicated that the 7 texture features of GLSZM, showed no significant difference over time, while the shape feature provided no additional information about the stability of the gel phantom. The varying stability of all first order and second order features measured might be underestimated. Since the study is ongoing, assessment of other pulse sequences results will be added.

This work establishes the acceptability of polymer gel phantom as a candidate for use as a radiomic feature stable phantom, when assessment of the identified texture features. Future studies will validate the findings with other pulse sequences.

Alginate hydrogels present potential for application in anthropomorphic tissue-mimicking phantoms due to their tuneable physical and magnetic properties [1]. Their capability to mimic tissue-relevant relaxation times makes them well-suited for applications in quantitative magnetic resonance imaging (qMRI), where inter-site reproducibility and consistency are essential for reliable biomarker validation and early cancer detection [2]. Despite their widespread use in biomedical applications [3], the long-term physical and functional stability of alginate-based phantoms needs to be fully characterized. This study aims to quantitatively assess the temporal stability of alginate breast phantoms by monitoring changes in MRI parameters, including T₁ and T₂ relaxation times, over an initial 28-day period. This work addresses the crucial need for standardized and reproducible qMRI calibration materials, especially in breast imaging applications, where reliable relaxation times reference values are needed for quantitative lesion assessment.

The prototype phantom consisted of twelve 15 mL Falcon tubes filled with 2% (w/v) alginate in saline (1.17% NaCl). To achieve target breast tissue relaxation times (Fig. 4) [4], two doping formulations were used: 0.42 mM NiCl₂ (for T₁ modulation) and 1.14 mM MnCl₂ (for T₂ contrast) to mimic fibroglandular tissue (FGT), and 3.10 mM NiCl₂ (T₁) and 1.13 mM MnCl₂ (T₂) to mimic adipose tissue characteristics. MRI data of the phantom (Fig. 1) were acquired on days 0, 14, and 28 using spin echo sequences (T₁: TR = 500–3500 ms, TE = 7 ms; T₂: TR = 500 ms, TE = 7–350 ms) on a 3 T Siemens Prisma. A freshly prepared phantom with the same composition was scanned at each timepoint to assess formulation reliability and temporal stability. Phantoms were equilibrated at scanner room temperature (19.2–20.2 °C) for at least 24 hours prior to imaging. Relaxation times were estimated using saturation recovery (T₁) and mono-exponential decay (T₂) models, with B₁ correction applied to account for RF inhomogeneity. Phantom reliability was assessed using the coefficient of variation (CV), whilst gel stability was evaluated using independent two-sample t-tests comparing days 0 and 28.

Phantom reliability was high across sessions (days 0, 14, and 28), with T₁ and T₂ measurements demonstrating strong longitudinal consistency for both FGT- and adipose-mimicking phantoms (Fig. 2), particularly in the FGT phantoms (CV = 0.28%). Regarding phantom stability, a significant increase in T₁ was observed in adipose-mimicking phantoms (p < 0.05), while no significant change was detected in the FGT phantoms (Fig. 3). T₂ values declined in both phantom types (Δ = −10.44 ± 15.68% and −9.83 ± 12.67%), though these changes were not statistically significant.

The phantoms demonstrated reliability, making them suitable tools for longitudinal qMRI assessment. The T₁ increase observed in adipose-mimicking phantoms likely reflects dopant redistribution or mild oxidation [5], whereas the greater stability (Δ = 6.66 ± 4.54%) of FGT-mimicking phantoms suggests that lower NiCl₂ content may inhibit such effects. T₂ reductions may reflect gel compaction or altered water mobility during aging [6], though high variability in standard deviation limits interpretation. The stable B₁ correction and narrow temperature range suggests that observed relaxation changes likely reflect intrinsic material behaviour rather than technical-induced variability. Despite relaxation times being systematically elevated (~13–15%) relative to literature values (Fig. 4) [3], their internal consistency across phantom types supports their use in comparative or calibration-focused qMRI studies. These findings highlight the potential of alginate-based phantoms as low-cost, tuneable models for advancing qMRI, by providing a means to assess the reliability of MRI relaxation times that mimic those of breast tissue. However, the relatively short 28-day monitoring period may underestimate long-term degradation, and variability in T₂ measurements suggests that future work should investigate strategies to improve gel uniformity and chemical stability. Additionally, exploring a broader range of dopants and storage conditions could further refine phantom stability and extend applicability to a wider set of imaging contexts.

This study confirms that alginate-based breast phantoms provide reproducible qMRI measurements over short timescales while revealing sensitivity to material aging, particularly in T₁ behaviour. These results underscore the need for periodic reassessment of phantom properties in longitudinal or multi-centre qMRI studies to ensure sustained measurement reliability. Continued refinement of phantom composition will be key to enhancing long-term stability for use in standardizing breast imaging protocols and quantitative biomarker development.

Evaluating and validating biophysical models is crucial for accurately characterizing tissue microstructure via diffusion MRI. Neurite Exchange Imaging (NEXI) is a two-compartment model of brain gray matter (GM) accounting for water exchange between isotropically oriented neurites and extracellular space [1-3]. NEXI is based on the analytical Kärger model of barrier-limited exchange between two Gaussian compartments. Although widely used, NEXI has not yet been evaluated in realistic numerical substrates of tortuous and beaded neurites, with varying levels of orientation dispersion and membrane permeability, which may challenge the Kärger assumptions. Indeed, previous simulation work either used the same analytical forward model for generating signals as is used for fitting the signals [1,2], or performed intra-cellular simulations in realistic substrates only [3]. Here, we assess NEXI's performance using Monte Carlo (MC) simulations of diffusion in realistic GM substrates with packed tortuous and beaded neurites.

Three substrates with orientation dispersion targets ☰ c₂ = 0.4, 0.6, and 0.8 were generated using the CATERPillar toolbox [4]. For each substrate neurites were grown in a (100 µm)³ voxel, using overlapping spheres with both beading (amplitude = 0.3 × initial radius) and tortuosity (ε = 0.4, standard deviation of the distribution of the 3D placement of the spheres) (Fig. 1). MC simulations used intra- and extracellular diffusivities of 2 and 1 µm²/ms, respectively; 0.8 µs step duration; 10⁵ random walkers; and no permeability. For the highly dispersed substrate (c₂ = 0.4), simulations were run with permeability values of 0, 10, 20, and 30 µm/s [5-7]. Each simulation was repeated 3 times for reproducibility. Diffusion-weighted signals were generated for 5 shells: 1 (12 dirs), 2 (16), 3.5 (24), 5 (30), and 7 (40) ms/µm², at diffusion times ∆=15, 26, and 38 ms, and diffusion gradient duration δ=4.5 ms to approximate the narrow pulse condition. T2 relaxation was ignored. NEXI was fitted using the Swiss Knife Toolbox (https://github.com/Mic-map/graymatter_swissknife) [1,2,8], fixing the exchange time (tex) to an arbitrarily high value in impermeable substrates. For each ∆, Diffusion Kurtosis tensor (DKI) (up to b=3 ms/µm²) and, for the impermeable substrates, the Standard Model of white matter (SMI) [9] were also fitted.

Resulting substrates had neurite volume fractions f=0.57, 0.56, and 0.55, and final c₂​ = 0.42, 0.64, and 0.73, matching the targets c₂​ = 0.4, 0.6, and 0.8, respectively (Fig. 1). Impermeable substrates of varying c₂: NEXI intra- (D¡,-41%) and extracellular diffusivities (De, -33%) were lower than the nominal (free) diffusivity, which is expected due to the effects of hindered diffusion in the extracellular space and along the neurites (irregularities from beading and tortuosity) (Fig. 2a). De was also lower in high-dispersion substrates as expected. The D¡ estimates were consistent with SMI estimates of intracellular diffusivity (Da) (Fig. 2c). NEXI slightly overestimated the intra-neurite volume fraction (f, +12%). Mean diffusivity (MD) and mean kurtosis (MK) showed some time-dependence in those substrates due neurite irregularities (Fig. 2b). Da from SMI (Fig. 2c) also displayed some time dependence, particularly for lower dispersion substrates. Permeable substrates: NEXI estimated tex between 10-31 ms (Fig. 3a), matching the theoretical expected exchange times given each permeability and cylinder radius [1,10]. Other NEXI parameters remained stable as a function of permeability. MK showed more marked time-dependence, increasing with permeability, while MD time-dependence was not affected (Fig. 3b).

Realistic neurite geometry yielded some time-dependence in MD, suggesting some contribution of structural disorder in this diffusion time range. SMI-derived Da scaled with 1/√∆, a hallmark of short-range disorder [11]. However, most time dependence in MK was found in the presence of permeability, consistent with the expected dominant influence of intercompartmental exchange [1,12]. Importantly, NEXI compartment diffusivity and neurite fraction estimates remained stable across permeability values, showcasing its ability to separate contributions from each compartment and their exchange. Finally, NEXI-derived exchange times aligned very well with the theoretical expectations.

By generating synthetic diffusion signals under controlled microstructural conditions, we evaluated the performance of NEXI to estimate key microstructural features of the GM brain tissue. Our findings suggest that NEXI can very reliably estimate exchange time across a range of environments with different permeability, as well as dissociate other model parameters from the exchange. Future work will focus on the influence of realistic noise levels and on the inclusion of cell bodies in the substrates.

Slow interactions between sodium nuclei and surrounding macromolecules generate a triple quantum (TQ) signal, which can serve as a biomarker for cell viability [1–3]. Agarose, a polysaccharide often used in phantom experiments, forms a solid gel after cooling, trapping water in pores ranging from sub-nanometer to hundreds of nanometers in size [4]. Low-concentration agarose induces sodium biexponential relaxation, on timescales like those in vivo, making it a suitable model system for the in vivo sodium TQ signal [5]. pH and temperature changes both alter sodium’s molecular environment. Whilst the influence of pH changes in protein-sodium interactions have been previously investigated [6], temperature and pH studies in the ubiquitously used agarose remain scarce. Hence, this study investigates the influence of pH and temperature on sodium transverse relaxation times and the TQ/SQ ratio in agarose samples.

Measurements were performed using a 9.4T preclinical MRI (Bruker Biospec 94/20) with a linear Bruker 1H/23Na volume coil. For the pH measurements, 154 mmol NaCl solutions with 12 different pH values were prepared by adding either 154 mmol NaOH or citric acid. The final pH was measured with a VOLTCRAFT PH-100 ATC device (Conrad Electronic SE, Germany). The final pH values were: 2.48, 3.62, 4.65, 5, 6.23, 8.05, 10.10, 11.27, 11.30, 11.56, 11.60, 11.89. A 2% w/v agarose phantom was prepared with the final solution. For the temperature measurements, 2% and 4% w/v agarose phantoms with 154 mmol NaCl were filled in a cylindrical flask (2.7 cm diameter, 4 cm height). The phantoms were heated using a flexible animal heat bath around the area for uniform heating, with temperature continuously monitored via a fiber-optic sensor in the center of the phantom. The signal was acquired after stabilizing the heat bath, with no temperature change exceeding 0.2°C within 5 minutes. Temperature was raised in 5 steps from 19°C to 43°C, then the phantom was allowed to cool and two subsequent measurements were performed. A TQ time-proportional phase increment (TQTPPI) pulse sequence was used to simultaneously measure SQ and TQ signals (Fiqure 1). The TQTPPI FID was non-linearly fitted using [5]: $$ Y(t) = \sin(\omega t + \phi_1) \cdot (A_{SQ,1}e^{-t/T_{2f}+A_{SQ,2}e^{-t/T_{2s}})+A_{TQ}\sin(3\omega t + \phi_2)\cdot (e^{-t/T_{2s}}-e^{-t/T_{2f}}) + DC, (1)$$ where $Y(t)$ is the TQTPPI FID amplitude, $A_{SQ,i}/A_{TQ}$ are the SQ and TQ amplitudes, respectively. $T_{2f}$ and $T_{2s}$ are the corresponding fast and slow transverse relaxation times. Sequence parameters were: $T_R = 400$ms, $\delta t_{evo}=200$us and $t_{mix}=125-135$us with a phase increment of 45° and 520 phase steps.

Fig. 2 shows the TQ/SQ ratio versus pH (left) and T2 relaxation times (right), with error bars representing the 95% confidence interval. The TQ/SQ ratio increased slightly at pH 5, plateaued at ~19% between pH 5 and 9, and rised steeply above 50% at higher pH. After a pH of 11.60 the curve started to flatten, suggesting sigmoidal behavior. The relaxation times showed an inverse trend. The slow component plateaued between 41–44 ms up to pH 10, while the fast component decreased from 16 to 9 ms by pH 10. Both dropped sharply after pH 10, reaching 0.37 ms and 20 ms at pH 11.89. Fig. 3 depicts the temperature dependence of the TQ/SQ ratio and relaxation rates. Both relaxation times correlate positively with temperature, confirmed by R²adj = 0.99 for all fits. The TQ/SQ ratio decreased similarly in both samples with temperature. After cooling, it was systematically lower than expected, unlike the relaxation times, which remained unchanged.

The behaviour of the TQ/SQ ratio in fig. 2 is driven by hydrogen bonding between hydroxyl groups [4]. At low pH, abundant +H ions inhibit bond formation, reducing negatively charged groups for ²³Na interaction and lowering the TQ/SQ ratio. Reduced bonding may also result in larger pores, decreasing the NaCl interaction surface and further lowering the ratio. At higher pH, more bonding enhances ²³Na-macromolecule interaction, raising the TQ/SQ ratio. The decreased relaxation times align with stronger bonding, which increases ²³Na interaction. For the temperature data, changes in relaxation times were likely caused by increased molecular motion’s influence on quadrupole interactions [7], not agarose structural changes, as its melting point is 65 °C. The systematic drop in TQ/SQ ratio after cooling, despite stable relaxation times, suggests possible sample changes detectable by this ratio, requiring further investigation, especially in protein samples.

The sodium TQ/SQ ratio exhibited strong pH dependence for values above 9, while remaining stable at physiological pH. Additionally, the study found a linear relationship between temperature and both relaxation times, as well as the TQ/SQ ratio. Both results confirm the sensitivity of the TQ signal on its environment and using the TQ/SQ ratio on is able to quantify this dependence.

Magnetic Resonance Imaging (MRI) presents a unique infection prevention and control (IPC) challenge due to its unique environment, complex workflows and distinct requirements for MRI-safe equipment [1]. Despite the clinical importance of IPC [2], specific guidance tailored to MRI remains limited. This study investigates the knowledge, attitudes, and practices (KAP) of MRI radiographers related to IPC for intravenous (IV) contrast administration, including the influence of training, workplace policies, and perceived infection risks.

An online cross-sectional survey targeting MRI radiographers in Australia was launched in October 2024. The survey gathered demographic information and evaluated KAP domains using Likert-scale and multiple-choice items. Additional questions focused on the challenges of IPC adherence, the availability of contrast-related IPC policies, the time since the last formal IPC training, and primary sources of IPC knowledge. Survey responses were collated and analysed using Microsoft Excel.

A total of 50 registered MRI radiographers completed the survey, 80% of whom were senior radiographers or MRI section leaders, and 76% had more than 10 years of experience in MRI. Study participants had high overall knowledge scores (93%) and positive attitudes (87%) toward IPC. However, only 66% always applied standard precautions when interacting with patients, with the remainder either frequently or occasionally applying standard precautions. Applying standard precautions was lower (52%) when utilising MRI equipment. Half (58%) of the participants had completed IPC training in the past year, while 46% and 14% either did not have or were unsure if they had an IPC team and IPC policies, respectively. Overall, in most of the participants’ workplaces, radiographers are responsible for setting the contrast injector, connecting the intravenous contrast to the patient, and cleaning the injector. However, only 26% of participants would clean the MRI equipment between patients, with 32% cleaning the room once a day. While institutional guidelines were commonly cited as the primary source of information, peer discussions were the most frequently reported source of knowledge regarding contrast injectors. Participants consistently identified MRI injectors, contrast tubing, and accessories as high-touch, high-risk equipment, with 70% believing their workplace recognised them as such. However, only 25% of participants would clean IV contrast-related equipment more than once per day. IPC adherence was reported to decline during high patient volumes, with respondents citing limited time to maintain aseptic technique during IV cannulation and contrast administration.

These findings indicate that MRI radiographers have a strong foundational knowledge and a proactive attitude toward IPC during contrast procedures. However, inconsistent workplace recognition of high-risk surfaces and reliance on informal learning may undermine their practice. Time pressure and the challenge of maintaining aseptic technique during IV cannulation, particularly in busy or emergency settings, further affect the consistency of IPC.

This study highlights a gap in MRI-specific IPC guidance, particularly regarding contrast administration. While radiographers demonstrate strong foundational knowledge and commitment to IPC, the lack of consistent, evidence-based protocols and modality-specific procedures undermines practice. Addressing these limitations requires a multifaceted approach, integrating formal training, MRI-appropriate IPC frameworks, and institutional support to promote consistent, safe, high-quality care. Radiographers are well-positioned to lead and inform these developments, and their active involvement in future research and policy design will be crucial to advancing IPC standards in MRI environments.

Simulation-based education enhances practical competencies ahead of clinical placements. In radiography, MRI simulation provides a safe, interactive environment for students to develop technical and decision-making skills. The addition of a scanning simulation tool was added in response to updated HCPC standards, which now require qualifying radiographers not just to assist but be fully capable of conducting an MRI examination. The aim of this study is to evaluate the initial experiences, perceptions, and acceptance of student radiographers in utilising an MRI simulator for their educational development.

A focus group with six final-year radiography students was conducted. Participants were self-selected volunteers who had completed an MRI simulation module. Discussions were transcribed and thematically analysed.

Six overarching themes emerged: Curriculum structure, Technical Challenges, Guidance and Support, Confidence and Skill -Building, Learning process, clinical practice. Some noted improved engagement in clinical discussions and job interviews. However, technical and instructional gaps were identified, with students requesting deeper explanations of MRI parameters. Students also suggested earlier and more frequent integration to better align with theoretical learning and clinical placements.

Students valued MRI simulation for developing competency and confidence but highlighted the need for earlier integration, improved usability, and stronger alignment with clinical practice.

Findings support structured MRI simulation implementation to optimise learning outcomes. Future research should explore the long-term impact on competency retention and employer perceptions of simulation-trained graduates.

The cartilage thickness measurement from MRI images has become a vital part of clinical trials and longitudinal OA studies. With the availability of automatic segmentation tools, the assessment of cartilage has become quicker and more accessible [1]. However, thickness measurement is still challenging, as the cartilage tissue spans only 2–10 voxels in-plane, and the inclusion or exclusion of voxels at the boundary can significantly affect the results [2]. Additionally, the measured thickness can vary with slice positioning, making 3D thickness measurement a potentially better option to the 2D approach. While consistent patient positioning during follow-up visits improves overall precision, it may not always be possible due to factors such as pain. The objective of this experiment was to assess the repeatability of thickness measurements when the knee is rotated and to compare the 2D and 3D approaches.

The left knees of eight healthy volunteers (4 men, 4 women, mean age: 35.5 ± 10.2 years) were scanned on 3T Siemens PrismaFit (Siemens Healthineers AG, Forchheim, Germany). The 3D DESS (TE=5ms, TR=14.1ms, 160 slices, 0.6x0.6x0.6mm3, flip angle=25°, acquisition=5:58min) was used. Each volunteer's patella center was marked with a black line, and two additional lines spaced 1cm apart were drawn on each side (Figure 1). A dedicated 15-channel knee coil was positioned to align these lines with the scanner laser, producing five distinct knee positions: neutral, two medial rotations, and two lateral rotations. Knee Position Angles were measured using RadiAnt DICOM Viewer (Medixant, Poznań, Poland) (Figure 2). Images were automatically segmented using MR ChondralHealth version 3.1 (Siemens Healthineers AG, Forchheim, Germany), with manual corrections applied when necessary. Nine femoral cartilage regions were assessed: medial (anterior/central/posterior), trochlear (lateral/central/medial), and lateral (anterior/central/posterior). Cartilage thickness was computed using a custom Python script. Cartilage voxels adjacent to the bone (proximal surface) were identified, and the real-world coordinates of voxel vertices shared between bone and cartilage at the bone-cartilage interface were calculated using the affine matrix. The distal surface voxel vertices were identified as vertices shared with the background (value 0) neighboring voxels. For each voxel along the distal surface, the nearest voxel along the proximal surface was found using a KDTree nearest-neighbor search (SciPy library [3]). Cartilage thickness was calculated as the mean of Euclidean distances between each distal surface voxel and its corresponding nearest neighbor on the proximal surface. This was performed either in 3D, or as 2D for each individual slice. The average thickness for each cartilage segment in 2D scenario was then obtained by averaging the thickness values across all slices within that segment. Slices with large discrepancies between the number of proximal and distal surface vertices, may result in overestimated cartilage thickness, therefore measurements exceeding 3 mm were excluded. Finally, correlations between knee rotation angles and thickness were evaluated using Repeated Measures Correlation calculated with R library rmcorr [4].

Results are summarised in Table 1 and Table 2.

Given the small number of probands and the large number of comparisons, the summary statistics presented in Table 1 should be interpreted as exploratory and primarily for orientation. The relationship between rotation angle and cartilage thickness varies across femoral regions and differs between 3D and 2D measurements. Notably, the trochlea medial segment shows a strong negative correlation in 3D, which is not observed in 2D. This suggests that precise and consistent positioning is essential when using 3D thickness measurements, particularly in longitudinal follow-up studies. Overall, cartilage thickness appears to be influenced by knee rotation in most femoral regions, with the exception of the medial posterior femur, where no correlation was observed.

This study demonstrates that femoral cartilage thickness is influenced by knee rotation, with region-specific patterns of correlation. The findings underscore the importance of consistent joint positioning in imaging protocols, especially for longitudinal assessments. Notably, the absence of correlation in the medial posterior region suggests localized variations in how rotation impacts cartilage morphology.

Parkinson’s disease (PD) diagnosis primarily depends on clinical criteria, but the need for objective biomarkers remains critical due to limitations in early and accurate detection. Neuromelanin-sensitive magnetic resonance imaging (NM-MRI) has emerged as a non-invasive imaging modality capable of visualising neuromelanin depletion within the substantia nigra pars compacta (SNpc), a hallmark of PD pathology. While initial studies suggest NM-MRI holds promise for enhancing diagnostic accuracy, considerable methodological variability presents a challenge to its integration into clinical practice. This systematic review focuses on evaluating NM-MRI’s diagnostic potential, examining technical variability, and considering its future role in PD diagnosis.

We systematically reviewed 29 diagnostic accuracy studies encompassing 1,207 patients with PD and 990 controls. Studies predominantly utilised 3T MRI scanners with T1-weighted fast spin-echo sequences, though considerable variation existed in protocols, image analysis techniques (manual, semi-automated, automated), and target anatomical regions within the SNpc. Diagnostic metrics- sensitivity, specificity, and area under the curve (AUC)- were extracted and rigorously analysed to determine NM-MRI's diagnostic performance.

NM-MRI consistently exhibited high diagnostic sensitivity across studies, with values ranging from 60% to 100%; notably, approximately one-third of studies reported sensitivities above 90%. Specificity varied more significantly, from 66.7% to 100%, with over 85% of studies reporting specificity exceeding 80%. Robust AUC values (0.80–0.99) were frequently reported, particularly when employing volumetric analyses of SNpc. Despite these strong diagnostic metrics in identifying PD vs healthy controls and even atypical parkinsonian syndromes (APS), significant methodological heterogeneity—including variations in MRI acquisition parameters, segmentation techniques, and clinical populations studied—contributed to observed inconsistencies in performance.

This systematic review evaluates NM-MRI's diagnostic utility for Parkinson’s Disease, highlighting strengths such as a high number of comparative studies and consistent use of validated diagnostic frameworks. However, methodological limitations like variability in imaging protocols and small-scale studies affect generalizability. Clinical implementation faces challenges due to the lack of standardized protocols; adopting T1-weighted fast spin-echo sequences and 3T MRI platforms is recommended. The SNpc region, especially its lateral part, is key for diagnosis, and combining analyses with the LC enhances accuracy. Workflow integration requires balancing scan duration and analysis methods, with automated tools showing promise. Standardization across technical parameters and analysis methods is critical for consistency. Although resource-intensive, NM-MRI could offer long-term cost savings through improved diagnostic accuracy. Future research should focus on achieving international protocol consensus, conducting large-scale validation studies, and integrating NM-MRI with AI and multimodal diagnostics for comprehensive PD management.

NM-MRI demonstrates robust potential as a sensitive and reliable biomarker for PD diagnosis and differential diagnosis. Addressing methodological variability through protocol standardisation and rigorous multicenter validation is essential to optimise its diagnostic accuracy, reproducibility, and integration into routine clinical practice.

Unilateral lumbosacral radicular pain is common, but accurately diagnosing it remains difficult [1]. Conventional macroscopic anatomical MRI (T1- and T2-weighted sagittal and axial sequences) produce a notable number of false positives or false negatives, resulting in discrepancies between MRI findings and patient symptoms [2-7]. In contrast, microstructural MRI, such as Diffusion Tensor Imaging (DTI), can reveal pathological changes in the form of reduced fractional anisotropy (FA) in the affected lumbosacral nerve, even in the absence of overt nerve compression [6-12]. However, there is no consensus yet on the optimal DTI protocol for reliably identifying symptomatic nerve roots [13]. In this experimental study, we aim to identify the most reliable DTI acquisition strategy for the robust visualization and segmentation of the lumbosacral nerve roots (LNR) (L3–S1), with the ultimate goal of enabling quantification of the extent of unilateral damage associated with radicular pain. We scanned healthy controls using multiple spinal cord DTI protocols and evaluated how well each protocol preserved anatomical normality. Additionally, we examined the effect of two distinct DTI processing pipelines to assess the relative impact of pipeline’s strategies on tractography results.

Seven healthy volunteers were scanned on a 3T Siemens Prisma system at the GIGA In Vivo Imaging platform, GIGA Institute, University of Liège, Belgium. Three different DTI protocols were considered (Table 1). ZOOMit showed higher anatomical detail but longer scan time (~35 min), while Coronal was fastest (~7 min). Due to time constraints, only two protocols were applied per subject, with each tested in four volunteers (Table 2). Two different DTI pipelines were examined to evaluate the impact of preprocessing on tractography: “SCT” and “our”. The former uses the MOCO algorithm [17], while the latter consists of denoising [19], correction for Gibbs ringing artifacts [20], correction of distortions due to susceptibility, head motion, and eddy currents [21]. “None” refers to a baseline case with no preprocessing. Tractography was performed using MRtrix3's tckgen algorithm [22]. We estimated FA using DTIFIT in FSL with weighted linear least squares [23]. The FA symmetry score was computed by multiplying the FA map and tract density map voxel-wise, then summing the weighted values separately for the left and right sides of the spinal cord (Figure 1).

Figure 1 shows left–right FA symmetry for the L3 to S1 nerve roots. Results are grouped by DTI protocol (Standard axial, ZOOMit axial, Coronal) and preprocessing pipeline (none, SCT, our). Substantial differences are observed across both protocols and pipelines. The combination of ZOOMit and our pipeline provided on average the highest FA symmetries for L3, L4, and L5. However, the Coronal DTI protocol combined with the our pipeline provided the highest FA symmetries for S1. Figure 2 shows representative tractography results of the lumbosacral nerve roots using the three examined DTI protocols. Notable differences in streamline density and anatomical coherence are observed across protocols, with ZOOMit axial DTI yielding the highest coverage but also more spurious tracts.

This study shows that the choice of DTI acquisition protocol has a substantial impact on tractography outcomes and left–right FA symmetry. The visibility and reconstruction quality of specific lumbosacral nerves varied across protocols: the ZOOMit axial DTI provided the most consistent and detailed tractography for the L3 nerve, while both the ZOOMit and Standard axial protocols performed comparably well for L4 and L5. In contrast, the S1 nerve was reconstructed with similar reliability across all three protocols. Notably, the ZOOMit protocol exhibited increased sensitivity to spurious streamlines, emphasizing the need for careful parameter tuning to minimize false positives. Between the two preprocessing pipelines evaluated, our in-house method yielded slightly better performance than the SCT pipeline. While this experimental analysis was conducted on a limited number of subjects, it offers preliminary results toward the optimization of DTI-based imaging of the lumbosacral nerve roots. Future studies involving larger cohorts are warranted to validate these findings and to further refine protocol selection for clinical and research applications.

In conclusion, our findings highlight that the choice of DTI acquisition protocol has a more significant impact on tractography and FA symmetry of the lumbosacral nerve roots than the choice of processing pipeline. ZOOMit provided the best anatomical detail for L3–L5, while all protocols performed similarly for S1. Careful tuning of tractography parameters and thoughtful pipeline selection remain essential.

Effective science communication is increasingly recognized as a critical component of responsible research practice, fostering transparency, enhancing public trust, and supporting reproducibility and collaboration across disciplines. Prior studies have shown that open access dissemination and targeted communication strategies can improve research rigor, democratize knowledge, and accelerate cross-border collaboration. In this spirit, the ESMRMB Early Career Researchers Committee launched the ESMRMB Podcast in 2020 to create an accessible and sustainable platform for the scientific community. This ongoing endeavor is entirely developed, hosted, and managed by early career researchers, with the overarching goals of enhancing knowledge exchange, fostering collaboration, and advocating for effective science communication. Additionally, the podcast aims to promote diversity in skillsets, academic backgrounds, career paths, and perspectives within the MR field. Here, we present a descriptive analysis of listener engagement, demographics, and global reach of the ESMRMB Podcast over a five-year period, highlighting the impact and value of this podcast for community-driven science communication.

A total of 16 episodes were recorded between October 2020 and April 2025 via Zoom with invited speakers, and edited using Apple iMovie. A total of 16 episodes were released, featuring content across a broad range of MR-relevant themes including MR technologies, clinical applications, career development, networking, grant writing, science communication, open access, international collaboration, and diversity, equity, and inclusion. The podcasts were distributed through major platforms such as Apple Podcasts and Spotify. Listener analytics were collected and evaluated over five years, including engagement trends, demographics, geographic distribution, device and platform usage.

Listenership remained consistent, with peaks typically aligning with new episode releases (Figure 1). The majority of listeners were aged 23–44 (87.3%) and identified as male (62.7%), reflecting strong engagement from early-career professionals and researchers (Figure 2). The podcast reached a global audience, with top listener countries including the United States (23.6%), United Kingdom (15.1%), Germany (14%), the Netherlands (8.3%), and Japan (7.9%) (Figure 3). Multi-platform access (Figure 4) and mobile-first consumption (57.6% via iPhone or Android devices) emphasized the importance of cross-platform accessibility in science communication.

Our podcast series demonstrated consistently high listener engagement, providing an open and accessible platform for scientific exchange across the global MR community. The high engagement among young listeners—particularly in countries like the US, UK, Germany, the Netherlands, and Japan—may reflect both a generational shift in media consumption for knowledge and skills improvement and the presence of well-established MR research centers that foster curiosity and ongoing learning. Moving forward, we envision enhancing our recording quality using professional audio equipment, exploring multilingual episode formats, and integrating podcast interviews with conference abstracts and featured presenters to further amplify scientific dialogs.

Our podcast demonstrates the potential of audio-based communication to enrich education, inclusion, and collaboration in the field of MR.

Artificial Intelligence, which can be used to solve tasks. Recently, there has been an exponential growth in publications using ML in medicine, but the quality of the articles is heterogenous. Thus, several quality scoring tools have been created over the years to assess quality of these articles using ML. The METhodological RadiomICs Score (METRICS) is one of them and it is the latest quality scoring tool to have been made available. Aim of this research is to perform a systematic review of the studies using radiomics and ML in neuroimaging of PPBT and to evaluate their methodological quality with the METRICS score.

Four radiologists performed systematic research of the scientific literature on Pubmed, Scopus and Web of Science including articles up to October 2024. Methological quality was assessed with the METRICS score [1].

In total, 361 articles were found, but only 51 of them were eligible considering the inclusion criteria, which were : population aged less than 19 years of age, original research articles, and use of radiomics, ML and/or Artificial Intelligence in neuroimaging of PPBT (fig. 1). The biggest part of the included articles was multicentric (41%) and used manual segmentation (58.8%). The METRICS score was moderate in 47% of cases with the highest score being 77 and 29.4 the lowest (fig. 2).

Our study has some limitations. Firstly, the studies included are retrospective. Secondly, MRI protocols are heterogeneous. Plus, some articles did not use the WHO 2016 classification to classify PPBT [2]. Finally, the METRICS score is relatedly new thus requires further validation.

ML has shown multiple promising applications in neuroimaging of PBT [3]. Although the overall quality of the articles in this field is moderate, it needs to be implemented to use these tools in clinical practice.

Sodium (23Na) MRI at ultra-high field (UHF) (B0 ≥7 T) enables non-invasive mapping of tissue sodium concentration (TSC), providing insights into cellular viability and ion homeostasis, with elevated sodium levels commonly observed in malignant tumours compared to normal or benign tissues [1][2][3]. However, various technical challenges remain—including low signal-to-noise ratio (SNR), rapid biexponential T₂ decay, and strong partial volume effects (PVEs)—hinder its clinical translation [4][5]. A variety of RF coil designs are available, such as surface coils, birdcage volume coils, single/dual-tuned coils, and phased array receive coils, with each offering different trade-offs in SNR and spatial resolution. This systematic review evaluates how different RF coil designs affect the diagnostic performance of quantitative ²³Na MRI based on clinical evidence.

A systematic review following PRISMA 2020 guidelines was conducted (Figure 1) [6]. PubMed was searched (2004–2024) using a structured string combining medical subject headings (MeSH) terms and keywords related to 23Na MRI and pathology. Studies were included if they quantitatively assessed TSC in clinical populations, reported means and standard deviations, and involved at least five participants per group. Extracted data included organ system, field strength, RF coil type, and TSC measurements. TSC contrast reflects the absolute difference in TSC between malignant and beningtissues, while effect size (d) is the standardised magnitude of the difference between separation of of malignant and bening tissue using TSC. Bar charts visualised performance differences across coil designs, grouped into surface coils, birdcage coil, knee-coil, clamshell coil, and more with respect to effect size (d) (Figure 3).

Fifteen studies were included, covering breast, kidney, prostate, and cardiac tissues, at various field strength (Figure 2). TSC contrasts ranged from 0.5 mM to 66.3 mM, with effect sizes ranged from 0.25 to 3.64. Clear separation (d > 0.8) was achieved in 9 studies, while others exhibited overlap between normal and pathological tissues (Table 1). For instance, Zaric et al. [7] demonstrated clear separation in breast cancer using a surface coil (d = 3.30), whereas Horvat-Menih et al. [8] and Barrett et al. [9] reported moderate separability in renal (d = 1.47) and prostate (d =1.09) imaging, respectively. Studies using multi-channel phased arrays for signal reception showed good performance with enhanced TSC contrast. Single-tuned sodium coils generally provided better performance than dual-tuned coils, while volume coils, such as birdcage, offered good B₁ homogeneity but lower sensitivity.

RF coil design was a dominant factor influencing the diagnostic accuracy of ²³Na MRI. The use of multi-channel phased arrays and/or single-tuned coils enhanced SNR and this leads to increased spatial resolution and improved tissue characterisation. Volume coils and dual-tuned coils exhibited lower sensitivity, which reduced the effectiveness of TSC-based tissue differentiation. Additional factors affecting coil performance included PVEs, tissue heterogeneity, and sequence optimisation [10]. These factors could be improved the use of independent proton coils for tissue localisation and B0 shim corrections. For sodium imaging, the use of volume transmit coils, which are particularly effective in achieving B₁⁺ homogeneity in large anatomical regions such as the torso combined with phased array receive coils to enhance sensitivity, will be essential for realising the diagnostic potential of ²³Na MRI in oncology, nephrology, and cardiology.

This review highlights that multi-channel phased surface arrays and the use single-tuned rather than double-tuned sodium coils significantly improve diagnostic performance in ²³Na MRI. Future RF coil developments could emphasis higher channel densities, flexible conformal surface arrays, advanced decoupling, and potential integration of volume Tx coil with phase Rx array in sodium combined with proton coil arrays, will further enhance clinical translation, supporting early disease detection and more personalised treatment strategies.

Type 2 Diabetes Mellitus (T2DM) is a prevalent and chronic metabolic disorder marked by high blood glucose level and insulin resistance. T2DM patients may experience cognitive impairments in areas such as executive function, memory, attention and visuospatial skills. Such cognitive changes are hypothesized to result from altered cerebral blood flow (CBF) and increased iron deposition in the brain. As such, advanced quantitative MRI techniques have emerged as essential tools to investigate the underlying neural mechanisms and potential biomarkers of cerebral dysfunction. This review focuses on studies involving the evaluation of the sensitivity and specificity of selected quantitative MRI techniques namely Arterial Spin Labelling (ASL), Quantitative Susceptibility Mapping (QSM), and T1/T2 relaxometry in detecting diabetes-related cerebral alterations.

Using PubMed, a systematic search was carried out to identify human studies in the English language involving investigations of the impact of T2DM on the brain through advanced quantitative MRI techniques. The search query used was: ("Diabetes Mellitus, Type 2"[MeSH] OR "Type 2 Diabetes Mellitus" OR "T2DM" OR "Type 2 Diabetes") AND ("Brain"[MeSH] OR "Cerebral" OR "Neuroimaging" OR "White Matter") AND ("Multiparametric Neuroimaging" OR "Quantitative MRI" OR "Quantitative Susceptibility Mapping" OR "QSM" OR "T1 mapping" OR "T2 mapping" OR "Arterial Spin Labelling" OR "ASL" OR "T1 relaxometry" OR "T2 relaxometry" OR "Relaxation time measurement") The earliest published paper from the search was from 2017. As a result, no strict date range was applied to the search period. Relevant earlier works cited in the included papers were also considered. The inclusion criteria required that studies: primarily investigate T2DM; include at least 10 participants per group; and utilize quantitative MRI methods. Reviews were excluded. The search resulted in14 research articles, which were subsequently analyzed using the PICO framework.

ASL-based perfusion imaging was the most frequently used MRI technique. However, studies assessing whole-brain CBF using solely ASL produced inconsistent results across different studies. To address this issue, ASL was often combined with fMRI, leading to more consistent insights particularly when calculating metrics such as the regional homogeneity (ReHo) to CBF ratio. Findings from these studies indicated region specific hypoperfusion in T2DM patients, notably in areas associated with memory and executive function, such as the hippocampus, precuneus, and posterior cingulate cortex. QSM was used in three studies to quantify brain iron deposition. Increased susceptibility values were observed in regions like the putamen, substantia nigra and Parahippocampal gyrus in T2DM patients. This was significantly visible on those with mild cognitive impairment. One multimodal study combined structural MRI, diffusion tensor imaging (DTI), ASL, resting-state fMRI, FDG-PET, and retinal OCT to assess the effect of T2DM on brain structure and function. The study found reduced attentional performance, lower nucleus acumens volume, and decreased cerebral glucose metabolism in individuals with T2DM, despite no significant differences in cerebral perfusion or white matter microstructure. Notably, no studies were found utilizing T1 and T2 relaxometry.

While CBF results derived from ASL alone were inconsistent, several studies emphasized the importance of multimodal MRI approaches in retrieving early signs of neural compromise in T2DM. Neurovascular Coupling (NVC) measures, derived from combining ASL with fMRI, showed stronger associations with cognitive performance and disease duration, offering a more detailed picture of how T2DM alters brain function. Notably, the integration of ASL with fMRI introduces non-quantitative measures, such as ReHo; nevertheless, this combination remains valuable for uncovering patterns of disrupted NVC. Studies also highlighted that voxel-based and region-specific analyses provided greater sensitivity than global measurements, uncovering localized abnormalities that whole-brain approaches might miss. QSM studies introduced iron deposition as an important mechanism of cognitive dysfunction, aligning with known pathways of oxidative stress and inflammation in T2DM. The relationship between elevated iron levels and impaired cognition, particularly executive dysfunction and memory loss, supports the use of QSM in monitoring disease progression and therapeutic response.

Quantitative MRI techniques especially ASL, fMRI, and QSM demonstrate significant promise in detecting cerebral alterations in T2DM patients. Multimodal imaging approaches and advanced analytical techniques such as voxel-wise and region-specific assessments enhance diagnostic sensitivity, potentially allowing earlier detection. Altered perfusion, iron deposition, and cognitive decline highlight neuroimaging biomarkers’ role in T2DM related brain dysfunction.

Magnetic Resonance Imaging (MRI) is a critical diagnostic modality that presents unique challenges for infection prevention and control (IPC) due to its strong magnetic field, non-ferromagnetic equipment requirements, and access-controlled environment [1]. A recent scoping review of IPC in medical imaging departments indicated that IPC education would benefit from a systems approach and further research [2].. The necessity of IPC within healthcare settings is universally acknowledged as critical to providing safe and high-quality care [3]. This scoping review aims to map the existing literature and identify the gaps in research concerning IPC practices within MRI. By examining the range and nature of studies, including guidelines, empirical research, and expert opinions, this review will provide a comprehensive overview of the current knowledge on the specialised IPC measures required in MRI suites. It will highlight the challenges and limitations of current practices around the world, outline the specific needs and considerations for effective IPC in MRI settings, and suggest areas for future research and policy development.

A scoping review methodology was employed in accordance with the Joanna Briggs Institute framework [4]. A comprehensive literature search was carried out utilising several electronic databases, including Medline, Scopus, CINAHL, Web of Science, and Google Scholar, to pinpoint relevant peer-reviewed articles. The search concentrated on topics pertaining to IPC within MRI settings. Searches were restricted to articles published in English from 2000 to 2024, with a focus on healthcare environments that utilise MRI and associated clinical workflows. Covidence was employed for data management and screening [5], while the SPIE [4] framework facilitated structured evidence synthesis. The PRISMA checklist was utilised to ensure transparent and efficient reporting [6]. Ultimately, a total of 1551 articles were identified; of these, 49 progressed to full-text review, and through further screening, 27 studies were included in the review.

The review encompassed 27 studies, predominantly concentrating on guidelines and recommendations for infection prevention and control (IPC) in medical imaging departments. A notable surge in published articles was observed in 2020. Of the 27 articles evaluated, 17 pertained to COVID-19 IPC, with only 4% specifically addressing MRI, primarily focusing on imaging-related services during the COVID-19 pandemic. Additionally, six articles were associated with investigative studies; one emphasised that patient contact items ranked among the most contaminated surfaces within MRI suites [7]. The four narrative reviews examined either hospitals or imaging departments that had established IPC guidelines, yet none specifically addressed IPC in the context of a magnetic field [2,8,9,10].

This review highlights the lack of MRI-specific studies in relation to IPC. Many studies mentioned MRI as part of the medical imaging department but there was a lack of specific interventions and guidelines dedicated to the MRI environment. The intricacies of MRI environments and the vulnerability of staff and patients to infections necessitate specialised IPC skills and practices tailored for radiographers and other personnel operating within these settings.

Despite the critical role of IPC in healthcare, there is a noticeable deficit in evidence-based guidelines specifically addressing the needs and challenges faced in MRI suites.

Functional MRI (fMRI) relies on blood oxygenation level-dependent (BOLD) signals, where neural activity–driven changes in oxygen consumption alter MR signal intensity. However, BOLD data are susceptible to neural and non-neural noise sources, head motion, and limitations of echo planar imaging [1]. Quality control (QC) is therefore needed to ensure data integrity and MR acquisition parameter consistency, particularly in multisite studies. This study developed a QC workflow combining qualitative and quantitative methods to assess resting-state (rs) fMRI data from a cohort of Swedish welders exposed to varying levels of manganese-containing welding fumes [2]. The aim was to use quantitative metrics to detect extreme outliers and to improve the overall data quality by minimizing exclusions.

Anatomical and rs-fMRI data were from the WELDFUMES study [2], involving 51 healthy male welders (ages 21–63) scanned at two sites using three 3T scanners: GE MR750w, GE Signa Premier, and Siemens Prisma. Imaging included T1-weighted fast gradient inversion recovery [2] and an EPI sequence (TR/TE: 1.5–2.5 s/30 ms; temporal resolution: 315–236 ms; voxel dimensions: 64×64 or 128×128) using 24", 48" and 64" channel head coils. Preprocessing of rs-fMRI and anatomical data, including realignment, unwarping [3, 4], and slice timing correction [5, 6], was performed using the CONN toolbox (RRID: SCR_009550, v22.v2407.a) [7,8] and SPM12 (RRID: SCR_007037, v12.7771) [9]. Volumes with frame-wise displacement >0.5 mm or BOLD signal >3 SD were classified as outliers [10, 11]. The outlier-excluded mean BOLD and anatomical data were normalized, co-registered into the standard MNI space, segmented into GM, WM, and CSF, smoothed (8 mm FWHM Gaussian kernel), and resampled (2 mm isotropic) [12, 13]. The default denoising pipeline in CONN [15] regressed out confounds (WM/CSF, motion parameters, and outlier scans), followed by high-bandpass filtering of BOLD signals above 0.01 Hz. Visual and automated QC was applied to the rs-fMRI and anatomical datasets: raw data QC assessed subject demographics, MR acquisition parameter consistency, and image quality; preprocessed data QC determined accuracy in segmentation, normalization, co-registration, and potential artifact; and denoised data QC evaluated the impact of residual noise on the distribution of functional connectivity. Sample-specific threshold values of Q3 + 3IQR or Q1 – 3IQR were used to identify extreme outliers [1].

One subject was excluded due to missing fMRI data. The fMRI sequence protocols differed in temporal resolution (TR = 1.5 or 2.5 s; 236 or 315 measurements) and spatial parameters (voxel sizes 3.4 x 3.4 x 3.7 mm³ or 3.8 × 3.8 × 3.6 mm³). Differences in RF sensitivity maps of the head coils (24", 48", and 64") may have contributed to signal variability in the BOLD data from the 48" and 64" coils, as shown in Fig. 1B. The effect of incorrect orientation on preprocessing is illustrated in Fig. 2. The results of the automated quantitative QC are shown in Fig. 3, where the metric Norm-Struct confirmed Subject 3's anatomical images as an extreme outlier due to incorrect slice orientation. In contrast, no extreme outliers were detected in the normalized (Norm-Func) and denoised (DOF) fMRI data; however, Subject 19 had 76 outlier scans, and a high MeanMotion was found in Subjects 19 (mean = 0.40) and 41 (0.41). Fig. 4 shows the distribution of functional connectivity (FC) and QC-FC correlations before (top) and after (bottom) denoising. FC was markedly biased and highly variable (mean correlation r = 0.23, SD = 0.305) but shifted closer to zero with reduced variability (r = 0.02, SD = 0.15) after denoising. The QC-FC association (r) had a mean = 0.03 (77% matches), SD = 0.13, which reduced to mean = 0.00 (88 % matches), SD = 0.16 after denoising.

In this study, we implemented an automated QC pipeline suitable for multisite MR data. Combining qualitative and quantitative methods, we assessed rs-fMRI data quality from the WELDFUMES project. Metrics such as Norm-Struct, Norm-Func, and visual inspection identified artifacts, acquisition inconsistencies, and evaluated key preprocessing steps including normalization, segmentation, and co-registration to MNI templates. Given the impact of motion and subject-specific factors on the BOLD signal, quantitative metrics such as MeanMotion, PVS, DOF, and FC distributions were used to improve the reliability of the fMRI data. QC-FC was then used to measure the association between functional connectivity and residual noise. Extreme outliers were defined using sample-specific thresholds, calculated as Q3 + 3 × IQR or Q1 – 3 × IQR [1]. A few subjects were in the extreme outlier zones (Norm-Struct, MeanMotion, and PVS, Fig. 3); however, there were no additional exclusions after denoising (DOF, Fig. 3).

In summary, the implemented QC workflow was highly effective at evaluating rs-fMRI data quality, and this approach could be adapted for task-based fMRI.

In medical imaging, accurate and clear visualization of diagnostic images is crucial for clinical decision-making, communication, and education. Magnetic Resonance Imaging (MRI) and similar modalities often rely on high-resolution images that may lose diagnostic value when displayed under poor lighting or with fixed contrast settings during presentations. This project addresses the limitations of static image enhancement by developing a real-time, user-controlled brightness and contrast adjustment tool integrated directly into Microsoft PowerPoint.

The developed tool utilizes Visual Basic for Applications (VBA) scripting within Microsoft PowerPoint, combined with Windows API functions, to allow users to adjust images of the current slide regarding brightness and contrast dynamically based on mouse movements and keyboard commands. The process begins when the user presses the "M" key, which triggers the monitoring. The program initializes the mouse position (startX and startY) and sets the variables isTracking = True and = False to start tracking mouse movements. The main loop then calculates the horizontal (deltaX) and vertical (deltaY) mouse movement to adjust brightness and contrast accordingly. As long as isPaused remains False, the adjustments continue. The brightness and contrast values are constrained within a range of 0 to 1. The program iterates through all shapes on the slide, checking if they are images (msoPicture). If so, the brightness and contrast are adjusted. Additionally, several key presses are monitored: "S" pauses the adjustments (isPaused = True), "R" resumes them (isPaused = False), and "X" resets all settings to their original values. Finally, the DoEvents function ensures the program responds to system events without freezing. The Program flow chart is shown in figure 2.

The tool enabled smooth, real-time image enhancement during PowerPoint presentations. It performed consistently across different medical image formats, including MRI, CT, and X-ray scans. Users were able to intuitively modify image properties mid-presentation, with immediate visual feedback. Enhanced visibility of fine details—such as tissue structure, lesions, or vascular patterns—was achieved, especially under suboptimal lighting. Presenters could adapt visual settings based on the audience and context, improving the clarity and effectiveness of medical communication.

The implementation proved to be robust and user-friendly. It required no prior image editing or additional software, making it practical for clinical, academic, and conference settings. The system allowed for dynamic, responsive control over image display without interrupting presentation flow. Feedback from test users highlighted ease of use, flexibility, and increased confidence in image-based explanations. The tool’s performance remained stable regardless of image resolution or presentation environment.

This project successfully introduced an intuitive, real-time brightness and contrast adjustment tool for PowerPoint-based medical image presentations. By leveraging native VBA capabilities and API functions, the solution improves diagnostic clarity without the need for external software. Future improvements could include window level/width adjustments for radiology applications and better support for high-resolution or compressed image formats. With continued refinement, the tool could become a standard feature for medical educators and clinicians aiming to enhance visual impact in digital presentations.

Throughout MR literature, often an ambiguity arises with the term “spoiled” or “spoiling”. In the generic field of gradient-echo sequences (mostly abbreviated as “GRE” or “FFE”)[1,2], “spoiling” usually refers to measures aimed to reduce unwanted transverse magnetization. However, the sequence can be “spoiled” (or sometimes “crushed”) according to – at least – three different variants: - RF-spoiled, which involves phase-modulating the excitation pulses[3] by n^2 ϕ_0, where n is the pulse number and ϕ_0 the golden angle or some other optimized angle, which may be 117 or 150 degrees. - Incoherent-gradient-spoiled, where a net gradient area of varying size is applied between successive RF pulses. Although the use of this type of spoiling is somewhat archaic, it cannot be ruled out that “spoiling” refers to this spoiling variant. Coherent gradient echo applies a constant nonzero gradient area between any pair of successive RF pulses. According to some definitions[1,3], this variant is not to be called “spoiled” at all, since it does preserve transverse coherence, as do balanced sequences. Nevertheless, many papers do call this choice “spoiled” (see an incomplete but representative literature review in the supplementary material). Many acronyms for spoiled sequences are available, but these can be ambiguous on their own, so many authors just publish their sequence as “spoiled”. It is the purpose of this abstract to quantify the occurrence of adequate specification of this term.

A set of publications were selected to examine the specification of the word “spoiling” (or “spoiled”). The was selection that was primarily drawn from references used in my PhD thesis and underlying publications[4–6]. A further selection criterion was the presence of the string “spoil” (as in spoiling, spoiled, spoiler) in the context of gradient-echo sequences. Hits in references were not considered. This delivered over 80 hits. By a random selection, the set was further limited to 50 hits. In these publications, the string “spoil” was examined on whether • It clearly specified the type of spoiling (e.g. by mentioning “RF-“, “Incoherent gradient-“ or “Coherent gradient-“ in the immediate vicinity of the word “spoiling”) or the immediate context specifies the type of spoiling. • It mentions “gradient spoiling”, without clarifying whether it is coherent or incoherent. • It is specified as “SPGR”, which may be RF-spoiling or incoherent gradient-spoiling. • Or it just mentions “spoiling”, leaving the reader to search through the context to guess which type of spoiling is used. The “immediate context” was defined as within one page of the first occurrence of the word “spoiling” (or “spoiled”).

Details of the list of publications can be seen in supplementary material (https://doi.org/10.5281/zenodo.15201724). The results are visualized in Figure 1.

Although the MRI-world is not really scarce on acronyms, there is no straightforward and unambiguous acronym for the coherent gradient-echo case. “FFE” (the Philips-entry in the MR-TIP[2] table) is clearly ambiguous, since “FFE” is also used as an umbrella-term for all types of gradient-echo sequences. Also “Spoiled FFE” is ambiguous, since it also can refer to RF-spoiled gradient-echo. “FISP” might be an unambiguous acronym, but this term is somewhat troubled by the presence of the technology called “trueFISP” (balanced, i.e. zero net gradient area between successive pulses). Is “FISP” then an umbrella term encompassing “trueFISP” as well as something like “untrue FISP”, or does “FISP” just refer to the latter? Is “FISP” referring exclusively to the “untrue” version of itself? “SSFP” (Steady-State Free Precession) might be another candidate for an unambiguous term for a sequence using coherent gradient spoiling. But it isn’t: although the one-time Shimadzu acronym referred indeed to coherent gradient spoiling, the MR-TIP table puts the GE-term “SSFP” in the same row as the Siemens-term “PSIF” (a Siemens-term referring specifically to the echo before the RF-pulse in a coherent gradient spoiling sequence); then it adds to the confusion by stating “A new family of steady state free precession sequences use a balanced gradient (…)” – which seems to make “SSFP” also an umbrella term. Such ambiguities might cause researchers to avoid acronyms altogether and specify their sequence as “spoiled”. Yet, as shown, 38% of publications insufficiently specify the type of spoiling.

When discussing gradient-echo MR sequences, the word “spoiled” (or “spoiling”) should not be used in isolation; these terms should be padded with either “RF”, or with “incoherent gradient-“, or with “coherent”. The latter (i.e. coherent spoiling, or coherent gradient-spoiling) should be read as “Gradient-echo sequence with a constant unbalanced gradient moment in each TR”. Of course, it is equally valid to not call this “spoiling” at all[1,3].

One of the hypotheses explaining the origin of Parkinson's disease (PD) is the abnormal deposition of iron in specific deep gray nuclei (DGN) [1]. In vivo quantification of iron from magnetic susceptibility (χ) maps obtained by quantitative susceptibility mapping (QSM) could provide an early neuromarker of PD. In this clinical context, optimization of acquisition parameters (number of echoes, spatial resolution, bandwidth) is important for robust χ quantification with reduced acquisition time [2].

Phantom: Tubes containing solutions of Dotarem® (Gadoterate Meglumine; 0.5 mmoL.ml-1; Guerbet; France) (Fig. 1) at various concentrations of known χ [3]. Healthy volunteers: Thirty-two healthy volunteers (HV) with no history of neurological disorder were recruited from 2021 to 2024 (Clinical trial NCT05107232; Univ. hospital of Rennes) and divided into two age groups (group 1: n=16; sex ratio=1; age=22±1 years and group 2: n=16; sex ratio=1; age=49±2 years). MRI data acquisition: QSM data were acquired at 3 T (Magnetom Prisma VE11C; Siemens Healthineers; Erlangen; Germany) using a 64-channels head coil. For phantom and HV, a 3D monopolar susceptibility weighted imaging (SWI) multi-gradient-recalled-echo (MGRE) sequence was used. Several parameters were evaluated such as the number of echoes, TR, TEmin/ΔTE, spatial resolution, and bandwidth. All acquisition parameters are resumed in Table 1. Acquisitions on phantom were repeated five times, during different MRI sessions. For HV, the protocol included 3D anatomical T1 and T2 weighted anatomical images (T1w and T2w) with a spatial resolution of (1 mm)3 to segment the DGN of interest. Image processing: Reconstructions of χ maps were computed on MATLAB (v2017Rb) using Sepia (v1.1.1). For phantom, the FSL brain extraction tool was manually settled to adapt it to the geometry of the tubes and three processing pipelines were evaluated [4]. The ImageJ (v1.53g; National Institute of Health) software was used to quantify the mean χ values and standard deviation (SD) of each manually traced ROI on χ maps. For HV, phase unwrapping, background field removal and field to susceptibility inversion were processed using an in-house developed processing pipeline based on ROMEO, PDF and MEDI algorithms [4] to calculate χ maps (Fig. 1, Fig. 2-A). In addition, six regions per hemisphere were segmented from the T1w image using the recon-all process of FreeSurfer [5] (v7.2.0) (caudate nucleus (CN), putamen (Put), globulus pallidus (GP)), and from the T2w image using pBrain [6] (substantia nigra (SN), red nucleus (RN) and subthalamic nucleus (STN)). The mean χ values and SD of each DGN were quantified using 3D Slicer (v5.6.2). Statistical analysis: For phantom, a non-parametric statistical test was performed to compare χ quantification between the acquisition parameters. For HV, a parametric statistical test was performed to compare χ values quantified in the DGNs between the two groups.

On phantom, increasing the spatial resolution involved an underestimation of the χ, due to the loss of signal-to-noise ratio, while reducing the number of echoes enabled reliable quantifications until five echoes. These observations led us to concentrate our in vivo acquisitions on the MGRE sequences up until five echoes. For HV, reducing the number of echoes from eight to five gave a quantification consistent with the literature [7] (Fig. 2-B), and no statistically significant difference was observed in function of the number of echoes used except in the group 2 for CN (8 vs 6 echoes and 8 vs 5 echoes: p<0.05). The aging effect was observed for the protocol using height, six and five echoes and was statistically significant for five DGN out of six (CN, Putamen, GP, RN: p<0.01; STN: p<0.05) demonstrating the robustness of a five-echo MGRE sequence in vivo.

In vitro, reducing the number of echoes from eight to five not impaired χ quantification. However, we demonstrated in vitro that a SWI MGRE sequence with only four echoes did not produce accurate and reproducible χ quantification within the ranges targeted in the DGN. In vivo, changing the spatial resolution from (1.5 mm)³ to (1 mm)³ not improved χ quantification. However, reducing the number of echoes to five allowed us to obtain a robust χ quantification consistent with the literature [7], while still enabling the detection of aging effect. These optimizations allowed us to reduce the acquisition time from 6mins31s to 3mins48s. The sequence is ready for inclusion in a multiparametric clinical protocol, with a spatial resolution of (1.5 mm)³.

These results make it possible to apply the MGRE with five echoes acquisition protocol on PD patients for in vivo quantification of χ in DGN, to explore the potential of this new neuromarker for monitoring PD progression and identifying various profiles of patients.

Lung MRI has long been constrained by intrinsically low proton density, short T2* relaxation times—limitation that is enhanced at high magnetic field strengths making it challenging [1]. Recent advances in ultra-short echo time (UTE) sequences and self-navigated respiratory gating now enable time-resolved, 4D pulmonary (3D + time) MRI in free-breathing conditions [2,3]. In this context, low field MRI systems (<0.55T) offer a trade-off between SNR and T2* relaxation [4]. Also, lower requirements for field homogeneities allow wider bores enhancing accessibility for claustrophobic or dyspneic patients. This study investigates the impact of magnetic field strength on image quality and dynamic pulmonary function assessment, using the exact same MR sequence. We hypothesize that the trade-offs inherent to field strength—particularly regarding signal-to-noise ratio and relaxation properties—can be quantified and exploited to optimize MRI-based lung imaging, with the potential of low-field systems for reproducible, radiation-free, region-specific imaging in free-breathing conditions.

Lung MRI was performed at 0.55T, 1.5T, and 3T (Free.Max, Sola, Skyra; Siemens-Healthineers) using a custom dynamic center-out radial UTE spoiled GRE sequence [5] with AZTEK trajectory sorting [3] and self-navigated respiratory gating. To ensure intra-subject comparability, the same healthy volunteer was scanned across three different clinical sites. The UTE sequence consists of a center-out radial acquisition lasting 11min with: TR/TE/flip angle=3ms/0.03ms/4°, FOV 32x32x30cm3, voxel size (2mm)3, matrix size [160,160,150]. The maximum gradient amplitude and slew rate were constrained by the lower hardware performance of the 0.55T system (15mT/m and 40T/m/s). Retrospective respiratory gating and image reconstruction were computed for 32 respiratory gates [6]. To assess the impact of magnetic field strength on image quality, respiratory-resolved images were evaluated based on the following criteria: the visibility and spatial extent of the pulmonary vascular tree and the contrast-to-noise ratio (CNR). CNR is computed as the ratio between lung-liver signal and the noise. To enable future inter-subject comparison, accurate lung segmentation is needed. Segmentation quality relies on sufficient CNR and clear delineation between lung parenchyma and adjacent muscular structures. Special attention was given to the diaphragm, whose substantial respiratory motion makes it especially challenging to segment consistently. The 4D MRI acquisition also provided full diaphragmatic coverage across the respiratory cycle, allowing dynamic segmentation required for registration and subsequent comparative analysis.

The visualization of the pulmonary vascular tree was assessed using MIP over 2 cm sagittal plane (Fig. 1). At both 0.55T and 1.5T, vessels were clearly visible up to the second generation, with well-defined contours and consistent parenchymal contrast. In comparison, at 3T, vascular structures appeared with lower contrast, and vessel contours were less defined. Qualitative assessment of the lung–diaphragm interface (Fig. 2) revealed notable differences in anatomical sharpness across field strengths. The signal profiles were sharply delineated at 1.5T, moderately at 0.55T and blurred at 3T. Quantitative analysis of CNR further supported these findings with the highest CNR observed at 1.5T CNR₁.₅=34, followed by 0.55T, CNR₀.₅₅=20, and lowest at 3T, CNR₃=4.9. To evaluate dynamic image consistency, respiratory-resolved images were analyzed across multiple gates (Fig. 3). The spatial distribution of normalized signal intensity remained stable across the respiratory cycle for all systems. However, minor motion blur was observed in gates corresponding to phases of high airflow particularly at 3T.

Our findings demonstrate that magnetic field strength has a direct impact on 4D lung MRI quality. While 1.5T provided the best compromise between anatomical sharpness and signal stability, 0.55T preserved essential structural details despite a lower CNR. The poor results at 3T are non-satisfactory and could be due to technical limitations. This can lead to a less effective retrospective gating resulting in lower SNR and under sampling artefacts after reconstruction [7]. Improvements are underway to reduce the noise of the 0.55T and refine the reconstruction of the 3T system.

In a travelling volunteer at 3T, 1.5T and 0.55T across three sites, 4D lung MRI was successfully reconstructed over 32 respiratory phases. Respiratory gating and anatomical delineation of the vascular tree and diaphragm were optimal at 1.5T and very encouraging at 0.55T. These findings support the potential of low and intermediate fields with ongoing developments aimed at optimizing low-field acquisition and reconstruction.

MRI data often lacks robustness due to variability introduced by site, scanner type, and hardware/software changes over time [1,2,3]. This variability can impact downstream analyses of brain structure, including measures of healthy tissue or lesion volumes [4,5]. The Cam-CAN dataset is a longitudinal, multi-modal, single-site study collected over 12 years, during which two major hardware changes occurred: a gradient coil replacement and a scanner upgrade (PRISMA vs. TRIO) [6]. We evaluated the robustness of Cam-CAN T1w structural data against these changes and whether pre- and post-processing techniques reduced effect sizes.

Two different datasets were used to evaluate hardware changes and FreeSurfer was used for cortical parcellation of 68 ROIs, with visual quality assessment and outlier detection. Inter- and intra- canner differences were assessed in a Traveling-Heads (n=20) repeated measures dataset of 5 patients each scanned twice on both TRIO and PRISMA within two weeks. Skull stripped images were pre-processed using three time-based (Z-Score, White-Stripe, Kernel Density Estimation, Fuzzy Clustering) and 2 sample-based (Nyul histogram matching and Least-Square Means) SI normalisation techniques. Intensity comparisons were conducted using normalized overlap (intersection), Kullback-Leibler (KL) divergence, and the standard deviation of subject-level normalized mean intensity (SD NMIs). Effects on ROI thickness measures were compared using Coefficient of Variance (CoV), ANOVA, and paired t-tests (with Bonferroni correction). The gradient-coil change was assessed on 89 matched subject pairs (matched on sex, handedness, and age) and assessed the efficacy of NeuroCombat harmonisation in mitigating these coil-related biases. We first quantified coil-related effects in the raw data using paired t-tests and Cohen’s d (with FDR correction) across ROIs, both unadjusted and age-adjusted via linear mixed-effects models (LMMs). NeuroCombat-harmonised outputs were then subjected to the same analyses. We compared harmonised vs. original values using (1) scatterplots with Pearson r and RMSE, (2) Bland–Altman plots (bias ±1.96 SD).

Scanner (TRIO vs PRISMA): Across all inter-scanner comparisons, the observed overlap values ranged from 0.84 to 0.98 and maximum KL divergence observed was 0.06 (approximately a percent similarity of 94%) which indicated high overlap between distributions. No inter-scanner ROI differences were detected (p<0.05). We examined intensity differences across scanners (PRISMA vs. TRIO) by calculating the standard deviation of subject-level normalized mean intensity (SD NMIs) across all four scanning sessions. Significant scanner-related differences (p<0.05) were observed in repeated measures ANOVA and paired t-test comparisons in up to 27% of ROIs depending on the normalisation technique used. Non-normalised images demonstrated scanner related differences in 9.7% of ROIs. Nyul and LSQ pre-processed images affected 5.5% of ROIs, while other pre-processing techniques demonstrated an increased scanner variability ranging from 13% to 16%. Gradient-coil change: Coil replacement induced moderate biases (Cohen’s d range ≈ –0.30 to +0.28; 12/82 ROIs significant at FDR < 0.05), persisting after age adjustment (β-coil range ≈ –0.27 to +0.25; 9 ROIs significant). NeuroCombat compressed effect‐size distributions to near zero (d range ≈ –0.02 to +0.02; no ROI significant post–FDR), whether unadjusted or age‐adjusted. There was a strong preservation of inter-subject variance (median Pearson r > 0.97; RMSE < 2% of mean volume across sampled ROIs), minimal systematic offsets on Bland–Altman (mean bias < 0.5% of mean ROI value; limits of agreement ±3%).

Scanner change introduce measurable shifts in raw intensities but had limited effects after FreeSurfer processing. The effects of pre-processing methods on ROI thickness measures were mixed, some introduced variability, while sample-based normalization techniques modestly improved scanner consistency. Although scanner-related differences were still detected in some analyses, their overall impact was limited. Gradient-coil–driven differences are not significant in FreeSurfer cortical thickness measures but NeuroCombat demonstrated a reduction in bias and compressed effect-size. These findings support its use in harmonising structural MRI data across scanner hardware changes in longitudinal and multi-site studies.

Limitations of our study included a small traveling-heads sample. Given, the small size, it would be useful to further explore of sample-based SI normalisation techniques to reduce scanner variability in FreeSurfer derived cortical thickness measures. Future work could test multi-site generalizability of these techniques in combination with ComBat. Overall, our findings support the robustness of the Cam-CAN dataset across major hardware changes but suggest that intensity pre-processing could further enhance inter-scanner consistency in structural MRI analyses.

Dementia encompasses a group of clinical syndromes characterized by progressive decline in cognitive functions, most often caused by neurodegenerative diseases such as Alzheimer’s disease (AD), frontotemporal dementia (FTD), or Lewy body dementia (LBD). [1] Differentiating between these etiologies, particularly in the prodromal or early symptomatic stages, remains a clinical challenge. Neuroimaging has become essential for differential diagnosis, using structural MRI or [18F]-FDG PET. However, FDG PET involves exposure to radiation and a high cost together with limited availability. Arterial spin labeling perfusion MRI (ASL-MRI) appears to be a promising biomarker for the assessment of neurodegenerative diseases, providing a quantitative, non-invasive measure of cerebral perfusion. [2] However, its clinical adoption is hindered by lower signal-to-noise ratio and sensitivity to artefacts, and need to be more explored before clinical use. This study aims to compare the diagnostic performance of multiple neuroimaging modalities—structural MRI, FDG-PET, and ASL-MRI—in differentiating major neurodegenerative dementia syndromes in a clinical setting.

The study population included 78 patients with typical AD, 11 with behavioral-variant FTD (bvFTD), 8 with semantic-variant FTD (svFTD), 13 with LBD, and 15 healthy controls (HC). Imaging data—including 3D T1-weighted MRI, ASL-MRI, and FDG-PET—were acquired using a single GE Healthcare Signa 3T PET/MR hybrid scanner. Two ASL sequences were employed in the study: a single-delay 3D pCASL sequence and a multi-delay sequence incorporating seven distinct post-labeling delays. Image preprocessing was conducted using SPM12 to obtain grey matter modulated probabilistic maps, FDG-PET scans and CBF maps in MNI space. Regions of interest (ROIs) were defined using 94 cortical regions from the AAL3 atlas [3], from which regional volumes, mean glucose metabolism, and mean cerebral blood flow were extracted. ASL sequence harmonization was performed at regional levels using the NeuroCombat algorithm. [4] Voxelwise group comparisons between patient subgroups and HC were conducted for each imaging modality using SPM12. Direct voxelwise comparisons between modalities were then carried out using W-score maps. [5] Covariates for all analyses included age, sex, total intracranial volume (for T1-MRI analyses) and sequence type (for ASL). Sparse Partial Least Squares Discriminant Analysis (sPLS-DA) was performed using the mixOmics R package [6] for three models: using regional volumes from T1-MRI alone (T1 model), combined with FDG-PET measures (T1+FDG-PET model), or with ASL measures (T1+ASL model) to classify AD, FTD (bvFTD and svFTD) and HC. Performance of the three models was evaluated using leave-one-out cross-validation and area under the curve (AUC) metrics to assess group discrimination.

In AD, atrophy was pronounced in the medial temporal regions, especially the hippocampus (pFWE<0.05), while FDG-PET showed bilateral hypometabolism in the parietal regions, posterior cingulate, precuneus, medial temporal areas, and in the dorso-prefrontal cortex. ASL revealed perfusion deficits in the parietal and posterior cingulate regions but with much less intensity and extent. In bvFTD, MRI and FDG-PET demonstrated predominant frontal lobe abnormalities (puncor<0.001), which were not consistently captured by ASL. In the case of svFTD, all identified modalities revealed left anterior temporal abnormalities, with ASL changes exhibiting less marked severity (puncor<0.001). In LBD, no significant atrophy was observed (puncor<0.001), but FDG-PET revealed hypometabolism in the occipital, parietal, and temporal lobes; ASL showed limited left parietal perfusion reduction. Intermodality discrepancies were observed across syndromes. In AD, greater limbic atrophy contrasted with broader parieto-temporal hypometabolism and hypoperfusion. In bvFTD and svFTD, structural and metabolic abnormalities exceeded perfusion deficits. In LBD, hypometabolism was more widespread compared to other modalities. The ability to discriminate between groups was enhanced by the addition of either FDG-PET or ASL to T1-weighted imaging. However, the discriminative performance of the T1+FDG-PET model was comparable to that of the T1+ASL model.

Despite less extensive abnormalities than FDG-PET, ASL showed similar spatial patterns and comparable classification performance, supporting its clinical relevance. [7-8] Limitations include diagnostic group imbalance, with AD predominance reflecting typical prevalence; ongoing recruitment aims to address this. Two ASL sequences were used, with statistical adjustments for quantification differences.

Combined with structural MRI, ASL offers diagnostic accuracy comparable to FDG-PET in differentiating neurodegenerative dementias, supporting its potential as a non-invasive alternative. Further recruitment will clarify its clinical value across syndromes.

Stroke remains a major contributor to global disability and death, ranking fourth in disability-adjusted life-years (DALYs) according to the Global Burden of Disease Study 2021 [1]. In Indonesia, the 2023 National Health Survey reported a stroke prevalence of 0.83% among individuals aged ≥15 years, reflecting a persistent public health burden despite a slight decline from previous years [2,3]. Ischemic stroke, caused by focal infarctions in cerebral, retinal, or spinal regions [4], can alter gray matter volume (GMV). GMV changes have been observed in patients with ischemic thalamic stroke [5] and with atrial fibrillation [6]. Diffusion-Weighted Imaging (DWI), a magnetic resonance imaging sequence sensitive to water diffusion [7], enables early infarction detection [8] and reveals GMV remodeling post-stroke [9]. DWI also enhances gray matter visualization due to its higher diffusion coefficient [7,10]. Given DWI’s accessibility in both tertiary and secondary care settings, this study aims to assess GMV alterations in ischemic stroke patients using a pragmatic, low-resource image processing workflow, with minimal computational dependency.

This study was conducted at the Radiology Department of Syaiful Anwar General Hospital, Malang, Indonesia, using 1.5 T and 3 T MRI scanners, RadiAnt DICOM Viewer, and MATLAB 2014b. The materials included ten axial DWI images from ischemic stroke patients and ten from healthy individuals. As shown in Figure 1, DICOM images were randomly selected and converted to JPEG. Parameters such as TR, TE, and b-values were analyzed. Image processing steps (Figure 2) included RGB-to-grayscale conversion, high-pass filtering, histogram matching, and median filtering to enhance contrast. Segmentation was performed to identify CSF, white matter, and gray matter using thresholding. The segmented image was converted to HSV to facilitate gray matter extraction through masking. Gray matter volume was then calculated and compared between stroke and normal groups to assess structural differences.

DWI acquisition revealed long TR values (2,700–5,900 ms), short TE (82–87 ms), and high b-values (1,000 s/mm²), suitable for tissue contrast and diffusion sensitivity. CSF appeared hypointense, while gray matter displayed hyperintensity due to its restricted diffusion. These imaging features were consistent across all patient and control datasets. Post-segmentation threshold values did not differ significantly between groups (p > 0.05). Mean GMV in ischemic stroke patients was 14.340 cm³, compared to 14.870 cm³ in healthy controls. Although stroke patients consistently showed reduced GMV across age groups, the difference was not statistically significant (p > 0.05). Variability in pixel spacing, slice thickness, and patient age likely contributed to the absence of significance. However, the segmentation pipeline successfully differentiated gray matter from other tissues in all samples, demonstrating reliability in a basic thresholding framework.

This study demonstrates a cost-effective DWI-based image processing workflow capable of extracting and quantifying gray matter volume in ischemic stroke patients. Although limited by a small sample size and the use of basic threshold-based segmentation, the approach remains viable in resource-limited hospital environments where advanced image processing infrastructure is unavailable. The imaging parameters used—long TR, short TE, and high b-values—align with established DWI protocols for optimal infarct visualization and diffusion sensitivity. While the non-significant GMV reduction may reflect infarction-related atrophy, such changes might be masked by demographic variability and differences in image resolution. Moreover, the lack of statistically significant findings should not diminish the practical importance of this pipeline, especially in developing settings. Importantly, this study highlights the feasibility of GMV estimation without reliance on atlas-based segmentation or machine learning algorithms, offering a starting point for further innovation. Future directions include expanding the sample size, standardizing voxel geometry, and integrating automated segmentation via convolutional neural networks to enhance diagnostic throughput.

Using DWI and a straightforward image processing method, this study demonstrates the potential to quantify gray matter volume in ischemic stroke patients, even in environments with limited computational resources. While no statistically significant difference in GMV was observed between stroke and control groups, the methodology reliably segmented gray matter and captured consistent volumetric trends associated with ischemic insult. These results reinforce the value of DWI as a versatile imaging modality for both diagnostic and research purposes. The developed workflow may serve as an initial framework for low-cost, quantitative stroke imaging and can be enhanced with automated tools in future studies for broader clinical use.

Cardiovascular diseases (CVD) are one of the leading causes of mortality and morbidity in patients with type 2 diabetes [1]. Those CVD, such as coronary artery disease or stroke [2–4], can arise before diabetes onset, at the prediabetic stage, with a greater risk for women [5]. Despite being an interesting therapeutic window due to its reversibility [6], first line treatments for prediabetes (lifestyle interventions) do not usually focus on the early cardiac alterations associated with prediabetes [7] and lack cardioprotective effects [8]. Investigating new sex dependent therapeutic strategies is then essential to limit cardiovascular complications in prediabetic women. Resveratrol (RSV) supplementation has previously shown cardioprotective effects on type 2 diabetic female rats [9]. We therefore aimed to compare its effects on the heart of prediabetic female rats to those induced by diet intervention, one of the conventional first line treatments.

40 female Wistar rats were divided into 4 groups fed for 5 months with a standard diet (CTRL), High-Fat-High-Sucrose (HFS) diet, HFS diet supplemented with RSV (1mg/kg/day in drinking water) during the last 2 months (RSV) or HFS diet for 3 months followed by 2 months of standard diet (RSD). We performed a longitudinal in vivo study of cardiac morphology, function and perfusion by magnetic resonance imaging at the 3rd and 5th months. Rats underwent an intraperitoneal glucose tolerance test to evaluate their prediabetic status. Finally, ex vivo experiments on isolated perfused hearts at 5 months were performed to simultaneously study cardiac function (Rate Pressure Product (RPP)) and energy metabolism (Phosphocreatine (PCr), ATP) with 31P magnetic resonance spectroscopy, during an ischemia-reperfusion injury (IR). Experimental protocol is shown in Figure 1.

After 3 months of HFS diet, increased myocardial wall thickness (MwT) and perfusion (p<0.01 vs CTRL) were found. HFS diet also induced elevated left ventricular end diastolic volume ((LVEDV) p<0.01 vs CTRL) and mass (LVM) (p<0.05 vs CTRL), along with glucose intolerance at the 5th month (p<0.05). HFS diet increased heart weight to tibia length ratio (HTLR) (p<0.01 vs CTRL), a known cardiac hypertrophy estimator [10]. HFS diet consumption was also associated with alteration of basal myocardial function and tolerance to IR, as shown by impaired RPP (p<0.001) and lower PCr and ATP during reperfusion (p<0.05, p<0.001 vs CTRL). Furthermore, HFS diet induced elevated end diastolic pressure (EDP) throughout the IR protocol (p<0.05 vs CTRL). RSV supplementation restored LVEDV, MwT and LVM (p<0.05 vs HFS), improved glucose tolerance (p<0.05 vs HFS) but exhibited no significant effect on myocardial perfusion. In addition, RSV restored HTLR to CTRL level (p<0.001 vs HFS) and improved myocardial tolerance to IR, characterized by higher RPP (p<0.05 vs HFS) with increased ATP and PCr during reperfusion (p<0.05 vs HFS). Interestingly, RSV also restored EDP over the course of the IR protocol duration (p<0.05 vs HFS). Alternatively, RSD exhibited no effect on LVEDV, MwT, LVM and perfusion impairments (p<0.01 vs CTRL) but restored glucose tolerance (p<0.001 vs HFS) at the 5th month. Furthermore, RSD failed to improve the elevated HTLR and EDP for the whole duration of the IR protocol (p<0.05 vs CTRL) but ameliorated tolerance to IR, characterized by increased RPP (p<0.05 vs HFS) and higher PCr and ATP during reperfusion (p<0.001, p<0.01 vs HFS).

5 months of HFS diet induced prediabetes, deleterious cardiac remodeling, increased myocardial perfusion but also reduced tolerance to IR injury, characterized by decreased myocardial function and PCr, ATP, as previously found in other prediabetic models [11–13] but also patients [14–16]. The two therapeutic approaches exhibited different cardioprotective effects here. RSV protected against deleterious cardiac remodeling induced by HFS diet in vivo and ex vivo but also improved glucose tolerance and myocardial susceptibility to IR injury. Interestingly, 2 months of a return to standard diet had no effect on deleterious cardiac remodeling induced by HFS feeding. However, RSD restored glucose tolerance to control levels but only improved myocardial susceptibility to IR injury. This observation could be explained in part by the lack of an effect on deleterious cardiac remodeling.

Interestingly, RSV, despite maintained HFS feeding, exhibited a stronger cardioprotective effect than 2 months of return to standard diet. Further studies are warranted to uncover the mechanisms involved, with a particular interest in mitochondrial function and oxidative stress. However, these results suggest that RSV supplementation could be an interesting approach to address the cardiac alterations associated with prediabetes, especially for women, which are at greater risk of CVD than men.

Cerebrovascular reactivity (CVR), the brain's ability to regulate blood flow in response to stimuli, is a critical indicator of cerebrovascular health. Impaired CVR has been shown in neurodegenerative diseases, such as Alzheimer’s [1], which has disproportionally higher prevalence in women [2]. Pregnancy represents a female-specific life event that leads to significant neuroanatomical changes, including in grey matter (GM), that remain years after baby delivery [3]. Women with a history of hypertension during pregnancy have been shown to exhibit reduced CVR [4] and increased dementia risk [5] compared to women with normotensive pregnancies, however the impact of pregnancy without hypertension on cerebrovascular health is not clear. In this work we investigate whether pregnancy history is associated with changes in global and regional CVR.

Multiple post labelling delay (multi-PLD) pCASL data (6 PLDs 250-1500ms, labelling duration 1400ms, 3.5x3.5x4.5mm³) from 40 female subjects was acquired on a 3T Siemens Prisma scanner during two conditions, normocapnia (6 min) and hypercapnia (5% CO2, 3 min). Six participants were excluded due to data quality issues. Analysis included two groups of women post-partum (14.9±5.8 months post-delivery): (1) normotensive pregnancy (n=13, 34.5±3.3 years), (2) hypertensive pregnancy (n=3, 33.3±2.1 years), and an age-matched group with 18 women that have never been pregnant (31.7±5.1 years). ASL data were analysed with FSL tools, including BASIL [6], yielding maps of CBF and arterial arrival time (AAT). Structural images were segmented to retrieve GM masks. Internal carotid artery velocities from phase contrast data were used to estimate labelling efficiency [7] and perfusion maps calibrated by hemisphere- and subject-specific labelling efficiency. CVR was calculated as the absolute change in perfusion over the change in end-tidal pCO2. Global and GM CVR and CBF metrics were compared using t-tests (p<0.05). A voxelwise analysis of CVR and resting CBF was performed using Randomise (threshold-free cluster enhancement, 5000 permutations) with age as covariate.

No statistically significant differences were found in either global or GM CVR between women with no history of pregnancy and women with a history of normotensive pregnancy (Figure 1). Similarly, no differences were detected in normo- or hypercapnic CBF between these groups (Figure 2). Voxelwise analysis did not identify any regions of statistical significance. Descriptively, CVR metrics were lower in the group of women with hypertensive pregnancy disorders (Figure 3), but statistical analysis was precluded by the limited sample size.

In contrast with previous reports of brain anatomical changes that persist years after baby delivery, cerebrovascular changes, evaluated through CBF and CVR, seem to be preserved following normotensive pregnancy. The preliminary observation of reduced CVR in women with history of hypertensive pregnancy disorders aligns with previous research and calls for further investigation in a larger cohort. It also highlights how CVR, as a functional measure, complements measures of CBF in assessing cerebrovascular function.

Using ASL MRI with hypercapnic challenge, we found no significant differences in global CBF and CVR between postpartum women with normotensive pregnancy history and women that have never been pregnant, suggesting preserved cerebrovascular function despite other research finding structural brain adaptations. These findings contribute to the growing body of research on the impact of sex-specific life events on brain health.

Extracorporeal circulation (ECC) is vital in cardiac surgery [1], yet its effects on flow dynamics and tissue perfusion are still debated. Comparing ECC with physiological blood flow is essential for understanding these effects. MR imaging of moving organs requires advanced tracking for reliable results [2]. In addition, the wide range of flow velocities during ECC present specific challenges for phase-contrast MRI. To explore the impact of the different perfusion types, we compared aortic flow during ECC with physiological conditions in a rabbit model using MRI. Non-pulsatile flow was measured using an MR-compatible ECC set up [3], while physiological flow was assessed using navigator-guided, retrospectively gated imaging, both in anesthetized rabbits.

All MR data were acquired from healthy rabbits using a 9.4 T Bruker BioSpec equipped with a Tx/Rx birdcage coil. Rabbits were examined under 3 perfusion conditions —physiological, antegrade, and retrograde—each tested on 7 rabbits. Flow MRI covered the entire aorta and major branching vessels, divided into thoracic, abdominal, and pelvic segments, achieving an isotropic spatial resolution of 0.5 mm. For the continuous ECC blood flow, a specialized MRI protocol utilized 3 Venc values (200, 50, and 20cm/s, TR/TE: 7.5/3.5ms) allowing for accurate reconstruction of a wide velocity range (Fig. 1). For physiological blood flow, Bruker’s FLOWMAP was adapted by adding a navigator, eliminating the need for complex physiological recording (TE/TR: 4/9.3ms, Venc: 100cm/s, 14 oversamplings), and retrospective gating was used to reconstruct pseudo-dynamic physiological flow (Fig. 2) [4]. All datasets were processed using ROMEO phase-unwrapping [5] and polynomial surface fitting on static tissue for background phase correction.

Both types of perfusion during ECC were conducted using similar volumetric inflow rates, with only minor variations between individual measurements. In the ascending aorta, significant differences in volume flow rates were noted between antegrade and retrograde perfusion, while antegrade perfusion and physiological perfusion yielded comparable results. In the biologically downstream segments of the aorta, including thoracic and abdominal areas, substantial differences in blood volume flow rates were observed among all three types of perfusion, with antegrade perfusion showing the highest volume flow rates. In the descending aorta, the volume flow rates for antegrade perfusion were three times higher than those for retrograde perfusion, while the value for physiological flow was intermediate between the two (Fig. 3). In the distal aorta, volume flow rates were similar for physiological and antegrade perfusion, whereas retrograde perfusion exhibited significantly higher rates. Conversely, in the supraaortic and visceral branches, no significant differences were found among the three types of perfusion.

Since EEC perfusion aims to replace cardiac function, it is preferable to achieve volume flow rates that are comparable to physiological conditions, even though pulsatility is absent. Evaluating perfusion conditions specific to the EEC scenario and comparing them to physiological conditions can provide valuable insights for developing optimal procedures for human applications. Using the triple VENC technique employed here, it was possible to accurately quantify different flow velocities, particularly in the inflow region of the cannula as well as in peripheral areas. As expected, due to the varying flow directions, differences in flow rates between antegrade and retrograde perfusion could be quantified. Interestingly, rather than following a strict pattern, differences between the two EEC methods and the physiological condition varied across different vascular regions. A significant challenge encountered was the segmentation of small peripheral vessels. For future studies, higher spatial resolution will be necessary to address this limitation.

Using a dedicated triple Venc approach for ECC and a navigator-based method for pseudo-dynamic physiological blood flow, different perfusion strategies could be systematically compared in a rabbit model. The presented methodology enables reliable evaluation of different ECC approaches in vivo, providing a robust foundation for future testing and optimization of new perfusion strategies intended for clinical application in humans.

Arterial Spin Labeling (ASL) is a non-invasive MRI technique widely used to measure cerebral perfusion. However, its application is often limited by substantial intra- and inter-subject variability, which complicates the differentiation between physiological and pathological changes [1]. Despite increasing interest in the female brain, the effects of hormonal contraception on cerebral perfusion and hemodynamics remain poorly understood. This study aims to address this gap by investigating the impact of hormonal contraception on cerebral perfusion (CBF) and arterial transit times (ATT) using ASL MR imaging.

MRI data in 9 women with a natural menstrual cycle (NC) and 15 women using hormonal contraception (HA, Desorelle 20) were obtained on a Siemens 3T Prisma Fit MRI scanner (64 channel head coil, UGent core facility Ghent Institute for functional and Metabolic Imaging (GIfMI)). Data were acquired at three (for NC group) and two (for HA group) timepoints during the menstrual cycle, for three cycles, resulting in nine datasets per volunteers. ASL was acquired using a multi-time point pulsed ASL (TR = 3500ms, labeling duration = 1.8s; post-labeling delay’s from 0.3s to 3.1s, increments 0.3s, 1 label-control pair), and processed using the Bayesian Inference for Arterial Labeling (BASIL) toolset [2] from FSL (distortion correction with field map, model with macrovascular component and spatial priors, voxelwise calibration). Regional CBF and ATT were quantified for nine brain regions based on the MNI152 atlas (caudate, cerebellum, frontal lobe, insula, occipital lobe, parietal lobe, putamen, temporal lobe, and thalamus). Additionally, blood samples were drawn to analyze concentration of sex hormones (estradiol (E2), progesterone (Prog), follicle stimulating (FSH), and luteinizing hormone (LH)). rCBF, ATT and sex hormones of both groups were compared using a Mann-Whitney test with a False Discovery Rate correction and the significance level of alpha = 0.05.

A significant lower level of LH (mean difference = -4.45 U/L, p<0.001), FSH (mean difference = -2.25 U/L, p-value < 0.05), E2 (mean difference = -45.40 ng/L, p-value < 0.001) and Prog (mean difference = -0.25 μg/L, p<0.001) was measured in the women using contraception compared to the women wih a natural menstrual cycle. Significantly higher perfusion in HA compared to NC was measured in all regions except the cerebellum, with differences ranging from 1.25 ml/100g/min (temporal lobe) to 14.09 ml/100g/min (thalamus) (Figure 1). Additionally, in HA, compared to NC, a significant higher ATT was measured in the cerebellum (mean difference = 0.031s, p<0.001), and lower in the frontal (mean difference = -0.019s, p<0.01) and parietal lobe (mean difference = -0.014s, p<0.05) and the thalamus (mean difference = -0.035 ml/100g/min, p<0.001)(Figure 2). No statistical significant differences in ATT were found in the other regions.

This study highlights the impact of hormonal contraception on cerebral perfusion and arterial transit time. HA users showed higher CBF in most brain regions, higher ATT in the cerebellum and lower in the frontal, parietal lobes, and thalamus. These findings emphasize the need to consider hormonal status in neuroimaging studies, as HA can influence CBF and ATT. Limitations include a small sample size, and future research should explore larger cohorts and various contraceptives.

Hormonal contraception use significantly impacts cerebral perfusion and ATT measurements, highlighting the need to account for hormonal status in perfusion imaging studies. These results contribute to a better understanding ofthe influence of sex hormones on cerebral hemodynamics.

Since its introduction about 15 years ago, Quiescent-Interval Single-Shot (QISS)1,2,3 has been widely used for peripheral MR angiography, e.g. for peripheral arterial disease (PAD)2,3. In a recent publication4, QISS was used to acquire high resolution hand angiograms using ECG triggering at 1.5 T. However, strenuous pre-scan preparation steps, prominent background signal and the presence of artefacts such as discontinuity and susceptibility were reported. Here, we ventured to address these issues by using a one-slab excitation, Peripheral Pulse Unit (PPU) out-of-phase5 imaging on a latest generation, high performance 3T system.

Healthy volunteers were scanned in prone position with one hand extended above the head (“superman” position) using a product 2D QISS sequence at 3 T with one or several “slab” adjustment volumes (120° slice selective excitation, Cartesian True-FISP (Trufi) readout, venous saturation band = 70 mm; MAGNETOM Cima.X, max. gradients / slew rate 200 mT/m(s), 16-ch. Hand-wrist coil, Siemens Healthineers). Two echo times, TE = 2.42 ms (in phase), and TE = 3.5 ms (out-of-phase) were chosen by setting the receiver bandwidth (BW) to 651 Hz/Px (protocol 1), 347 Hz/Px (protocol 2, full echo) and 195 Hz/Px (protocol 3, 2/3 partial echo), respectively. Note that the sequence didn’t allow for setting the TE separately yet. Other parameters were: Quiscent-interval TI = 345 ms, Peripheral Pulse Unit (PPU) triggering, Single 3D slab reconstruction, TR = 687.61ms, interpolated voxel size = 0.2x0.2x1.0 mm3, matrix size= 640X416p, 256 slices, FOV = 165x108 mm2. Contrast to Noise Ratio (CNR) was estimated in three arteries: Ulnar, Radial, Proper Palmar Digital arteries in three distinct imaging slices using: CNR=(S1−S2)/σ S1 is the mean signal of a region of interest (ROI) in an artery in a single slice, S2 is the mean signal of a ROI of background tissue BT, σ is the standard deviation of the noise outside the object.

Using a single slab averted all block-like artefacts reported before (Fig. 1), 4 and resulted in good quality arterial QISS angiograms in short scan times of 3:36 min (Fig. 2). Using the PPU instead of the ECG for triggering did not deteriorate the apparent image quality; the arteries of the hand were well visualized with no additional artefacts. Using the same out-of-phase-TE, but increasing the bandwidth, reduced the tissue signal by signal (BT, Protocol 3) / signal (BT, Protocol 2) = 1.2 (Fig. 4). Of course, these results are affected by the echo time and bandwidth – see discussion. For all protocols, the maximum intensity projections (MIP) of the data exhibited similar arterial structures (Fig. 2). Susceptibility artefacts manifested as smearing of the arteries more prominently at lower bandwidths, discontinuities were present in one location at BW = 195 (Fig 2c). The CNR has increased for smaller bandwidths and for out-of-phase TE across all three arteries (Fig. 3), except for the radial artery in volunteer 2 (in color red Fig. 3).

Using a single instead of multiple adjustment slabs resulted in MIPs free of the block artefact (Fig. 1), likely because more uniform adjustment parameters were used. PPU has provided good results for hand angiography, although previous studies1,2,4 deemed PPU to be inferior. The difference may be caused by faster PPU electronics of the system used. Setting TE to cause fat and water spins to be out-of-phase is an established method to reduce lipid signals (TE = 1.23 ms, 3.69 ms, 6,15 ms, 8.61 ms, …. at 3T). When we set TE close to the out-of-phase condition, the background tissue around the arteries is noticeably darker, indicating better suppression of the static tissue, and consequently, enhancing the subjectively observable contrast between the arteries and the static tissue (Fig.1b and 1c, CNR in Fig. 2). For all three arteries, CNR was higher for TE= 3.4 ms/ BW = 195Hz/Px. This imaging protocol have yielded a near full visualization of the major hand arteries without requiring the lengthy pre-scan preparation set-up procedures (moisturizing hands, using clay mold)4. Of course, the receiver bandwidth has an effect on CNR, too. Higher bandwidths are associated with less signal and higher noise due to the higher sampling rate. This fact skews the comparison of CNR in favor of data with lower bandwidths (Fig. 3). On the other hand, lower bandwidths risk exacerbating susceptibility artifacts (Fig. 2). We are currently implementing a sequence that allows for setting TE individually. A high bandwidth, out-of-phase TE sequence appears to be ideal to minimize artifacts and maximize static tissue suppression.

The developed protocols, relying on one-slab-adjustments, PPU triggering and out-of-phase imaging, provided high quality angiograms of the human hand without contrast agents, and averted the issues reported before. More experiments are needed to isolate the individual effects.

It is known that atrial fibrillation (AF) and chronic kidney disease (CKD) are common conditions which are closely related and usually coexist. Each disease influences progression of the other. However, it is not well understood yet how AF affects renal blood flow. 2D phase-contrast magnetic resonance imaging (2D PC MRI) is a non-invasive method capable providing the functional information about renal artery blood flow (RABF) and quantify it. Therefore, we utilized 2D PC MRI to compare RABF in the same patient during AF and consequently in sinus rhythm (SR) after catheter ablation (CA) in general anaesthesia.

This pilot study enrolled 7 patients (age: 62 ± 13 years, F/M: 2/5) who underwent CA for AF in general anaesthesia. All subjects provided written informed consent with the participation in the study. The study was conducted in compliance with the principles of the Declaration of Helsinki and with the approval of local ethics committee. All the patients underwent MR examination 24 hours before CA and consequently 24 hours after CA. The examination was performed in the supine position during breath-hold exhalation using 3T VIDA MR system (Siemens Healthineers, Germany) equipped with 30-channel surface coil and 32-channel spine coil. MRI protocol included T2 TruFI sequences in 3 orthogonal orientations (thickness: 3 mm) for renal anatomy visualization and 2 sets of 2D PC MR sequences (1. FOV 250x250 mm, flow-encoded velocity range (VENC) = 100 cm/s, TR/TE = 40.80/2.82 ms; 2. FOV 200x200 mm, VENC = 80 cm/s, TR/TE = 40.64/2.99 ms; NA = 1, flip angle = 20°). The flow data were obtained in a plane perpendicular to the right and left renal artery (RA), approximately 10 mm from the ostium. 2D PC MR measurements were repeated 2 or 3 times to precise RABF quantification. Retrospective ECG or pulse triggering with RR-based arrhythmia rejection was used. All the data sets were normally distributed. Paired t-tests were used for statistic evaluation of the changes in patients before and after CA. P-value < 0.05 was considered statistically significant.

All patients presented with normal SR after CA. No significant differences in renal function parameters were found between AF and SR (creatinine before CA: 94±22 µmol/l, after CA: 97±25 µmol/l; cystatin C before CA: 1.1±0.3 mg/l, after CA: 1.1±0.3 mg/l). Paired t-test showed significantly decreased cross-section area (ROIs) of both renal arteries, increased RABF and increased renal blood peak velocity in each patient with SR after CA compared to the state in AF before CA (Figure 1, Table 1).

2D MR PC measurements were repeated 2 or 3 times for each FOV and VENC assessment to precise selection of RA cross-section area. Considering the size of the RA caliber and the realistically achievable spatial resolution when measuring RABF, it was not possible to reliably use automatic ROI delineation. RA ROIs needed to be selected manually which increases the probability of bias selection. The relative error for all evaluated parameters ranged from 10 – 20 %. Although the difference in RA cross-section areas before and after CA may have been biased by both ROIs selection error and renal artery wall pulsatility, neither RABF nor renal blood peak velocity can be influenced by this difference.

Restoration of SR following CA of AF is associated with an increased renal blood peak velocity and higher renal blood flow. These preliminary results require further investigation and verification in future studies. Acknowledgements This study was supported by the project National Institute for Research of Metabolic and Cardiovascular Diseases (Programme EXCELES, Project No. LX22NPO5104) – Funded by the European Union – Next Generation EU. This work was also funded by the project (Ministry of Health, Czech Republic) for development of research organization 00023001 (IKEM, Prague, Czech Republic) – Institutional support.

Carotid atherosclerosis remains a major risk factor for stroke in western countries[1]. While clinical management typically depends on stenosis severity, stenosis alone often fails to predict stroke risk accurately [2,3]. Treatment decision is currently based on the severity of stenosis, even though, clinical trials have shown that stenosis alone may not reliably predict cardiovascular risk in individual patients. Plaque composition and compromised endothelial function (in response to vasodilating stressors) are crucial in plaque development and the occurrence of cardiovascular events[4–6]. Consequently, an imaging method to detect features of high-risk plaque and luminal stenosis simultaneously would have added value in patient diagnosis. The recently proposed BOOST (Bright- and Black-blOOd phase-SensiTive) sequence offers simultaneous, co-registered, contrast-free bright- and black-blood imaging in a single scan, enabling both lumen and vessel wall visualisation[7]. BOOST has been previously used at 1.5T to identify high-risk coronary plaque features (thrombus and intraplaque haemorrhage)[8,9] and for anatomical assessment of the aorta[10]. However, the application of BOOST at 3T for 3D high-resolution imaging of carotid plaques has not been explored. Here, we optimised the iT2prep-BOOST sequence for carotid vessel wall imaging at a 3T scanner. This study was designed to selectively identify patients with high-risk carotid atherosclerosis by detecting compositional features of high-risk plaques and assessing endothelial (dys)function in a fast, high-resolution, contrast-free MRI session.

Building on our previous work [7], we have implemented an accelerated, 3D free-breathing, motion-corrected iT2 prep-BOOST research sequence on a 3T MR system (MAGNETOM Vida, Siemens Healthineers AG, Forchheim, Germany) for carotid artery imaging (Fig. 1). iT2 prep-BOOST allows for simultaneous bright and dark-blood imaging Images were acquired using a 64-channel head and neck matrix coil (Head/Neck 64, Siemens Healthcare) in 4 healthy subjects (female; mean age 34±14) with cardiac synchronisation via peripheral pulse sensor unit to optimise the acquisition parameters (data not shown). Patients undergoing carotid endarterectomy (CEA) (n=6) were recruited (Fig.2A). The iT2 prep-BOOST sequence was used for carotid plaque imaging at high resolution (0.9 x 0.9mm) and speed (~6 min). Endothelial (dys)function was measured using a phase-contrast MRI protocol before and 120 seconds after a cold-pressor test (CPT). Plaque-to-muscle ratio (PMR=SI plaque/SI reference muscle) using the black-blood images and changes in vessel area and blood flow were analysed. Tissue was collected postoperatively to validate the imaging data.

Results of the proposed simultaneous lumen and vessel wall BOOST sequence are shown for three patients. MR angiography and examples of bright and black-blood images acquired with iT2 prep-BOOST, show excellent delineation of the carotid artery lumen and vessel wall in all patients (Fig. 2B). The PMR measured on MRI was higher in plaques presenting with intraplaque haemorrhage as verified by en face tissue inspection (Fig. 2C). Assessment of endothelial vasomotor response to CPT showed vasocontraction in all three cases (negative % change in area) and reduction or no change in blood flow (negative or close to zero).

This study is the first to demonstrate the feasibility of a rapid, high-resolution MRI approach for simultaneous assessment of high-risk features in carotid plaques and endothelial function under 20 minutes. The iT2-prep-BOOST sequence provided clear delineation of the arterial lumen and vessel wall, and showed a higher plaque-to-muscle ratio in high-risk plaques with intraplaque haemorrhage. Cold pressor testing showed endothelial dysfunction characterised by no changes or reduced vessel area and blood flow at sites of plaque. This integrated approach offers a time-efficient method enabling the assessment of both plaque composition and endothelial cell function in a single MRI session, with the potential to improve risk stratification of carotid atherosclerosis.

In this translational study, we demonstrate for the first time that in vivo imaging of both high-risk plaque features and endothelial (dys)function can be assessed in a single MRI session at high resolution and in under 20 minutes. Such a strategy that provides both anatomical and functional assessment of the plaque may enable better stratification of atherosclerosis and selection of patients at risk.

Cerebral Autosomal Dominant Arteriopathy with Subcortical Infarcts and Leukoencephalopathy (CADASIL) is the most frequent hereditary cerebral small vessel disease (SVD) worldwide. The condition caused by characteristic mutations of the NOTCH3 gene leads to recurrent stroke, motor impairment, gait disturbances and cognitive decline [1]. Common magnetic resonance imaging (MRI) findings include white matter hyperintensities on T2-weighted images, lacunes and brain atrophy. Cerebral microbleeds (CMBs) described as small hypointense areas related to focal iron deposits on MRI sequences are observed in about one third of patients [2]. The predictors of CMBs in CADASIL are unclear, as the exact cause of their emergence and growing number remain poorly understood. Analyzing CMBs is however challenging for two main reasons: 1) CMB count represents a discrete and typically zero-inflated variable with a long-tailed distribution [3]; 2) CMB counting becomes less reliable when their number increases (see colleague submitted abstract [4]). In the present study, we used a quantitative and semi-quantitative approach to assess CMBs in CADASIL and sought to compare differences between CMB predictors obtained using two different analytical strategies.

We applied a two-step modeling framework to data obtained from 517 CADASIL patients recruited at the National Referral Center in France. Key variables were considered in addition to CMBs for analysis such as age, sex, vascular risk factors, the NOTCH3 mutation location, and key biological and neuroimaging information. In the first approach, the presence of CMBs was modeled using binary logistic regression (Logit) and thereafter CMB count (≥1) using a truncated negative binomial regression (NB). In the second approach, a binary logistic regression was also used at first, and thereafter an ordinal logistic regression (OLogit) based on four number categories. Continuous variables were standardized. Missing data were handled via mean imputation (imaging results) or via multiple imputation by chained equations (other data). Covariates were first screened for multicollinearity. A univariate analysis was then used to exclude those unlikely related to CMBs (p > 0.3). Age, hypertension, other neuroimaging measures and mutation location were retained in all models. The final model predictors were selected using stepAIC (OLogit) or LASSO (Logit and NB). When possible, model performance was assessed using a 70/30 train-test split. Model fit was evaluated with log-likelihood, Akaike Information Criterion (AIC), and Bayesian Information Criterion (BIC), while predictive performance was assessed using confusion matrices and mean absolute error (MAE). All analyses were conducted in R Statistical Software version 4.4.3 [5].

For both approaches, the Logit model identified age, hypertension, the number of lacunes, and diastolic blood pressure as predictors of CMB occurrence (Table 1). Using quantitative data, the NB model showed that brain parenchymal fraction, number of lacunes, MRI sequence, hemoglobin level and both systolic and diastolic blood pressure predicted the number of CMBs (Table 2). In contrast, when CMB were considered by number categories, the OLogit model identified the number of lacunes as the unique predictor (Table 3). In terms of predictive performance, the Logit model demonstrated strong classification ability, with high sensitivity and a robust F1 score, though it tended to overpredict the presence of CMB. The NB model exhibited moderate predictive accuracy, underestimating higher CMB counts, likely due to challenges in modeling overdispersion. The OLogit model showed limited capability in correctly classifying higher CMB categories.

Our two analytical approaches yield different sets of CMBs predictors in CADASIL. The NB and OLogit models are not comparable. Choosing a quantitative or semi-quantitative measures of CMBs leading to different analytical procedures appears thus crucial and should align with the exact underlying biological question. If the goal is to identify all predictors of CMBs, the first approach should be adopted despite the lack of precision in higher counts. If the goal is to obtain an accurate prediction, the second approach is preferable at the cost of losing information. Limitations of this study include the moderate sample size and CMB detection measures. Additional analysis of longitudinal data should help improve predicting CMB in CADASIL.

The results of this study further illustrate that the decision to quantify a biomarker such as CMBs should be based primarily on the ultimate objectives sought. Analysis of such data leads to a trade-off between the desired predictive performance, the complexity of the model to be developed, and the type of sample to be analyzed. In the future, enriching the information with longitudinal MRI data should further improve the prediction of CMBs in CADASIL.

Cerebral Autosomal Dominant Arteriopathy with Subcortical Infarcts and Leukoencephalopathy (CADASIL) is a monogenic cerebral small vessel disease (cSVD) caused by stereotyped cysteine mutations of the NOTCH3 gene encoding a receptor of smooth muscle cells and pericytes of arterioles and capillaries. This condition may lead to stroke, motor impairment, gait disturbances and cognitive decline [1]. On cerebral MRI, several MRI markers are observed during the disease course, including cerebral microbleeds (CMBs). CMBs are small hypointense dots on T2* and SWI sequences related to focal iron deposits (see Fig1). They are observed in 1/3 of patients [2]. Because of the higher sensitivity of SWI, the number of CMBs may be found higher than on T2*. As the individual identification of CMBs may be tedious both visually and quantitatively, some researchers labelled CMBs as ‘certain’ or ‘uncertain’, and others proposed to categorize their number in different classes [3–7]. As the method used to analyze CMBs may largely impact the identification of the potential predictors of such lesions (see companion submitted abstract [8]), we here assessed how the number of CMBs, the method used for counting and the type of MRI sequence can influence the observer agreement in repeated and independent assessments.

Thirty T2* and thirty SWI MRI were acquired respectively on a 1.5T GE scanner or on a 3T SIEMENS machine in CADASIL patients included in the National French referral center for rare vascular diseases of the eyes and brain (https://www.cervco.fr). On each scan, CMBs were visually identified twice by a trained scientist (Rater1 performed Reviews 1 and 2), and once by an expert neurologist (Rater2). Counts were then transformed in CMBs categories according to the CADAMRIT instrument (CADAMRITlike items) [7]. The continuous and categorial agreements were evaluated with the ICC3 and weighted kappa scores respectively [9,10]. A linear regression and a category-based analysis were performed to evaluate the influence of the burden of CMBs onto the results. For each test, the MRI sequence was considered as a factor that could potentially influence the results.

In the T2* sub-dataset, 63% of individuals were women, the mean age was 60.22 +/- 9.04 years. In the SWI sub-dataset, 50% of individuals were women, the mean age was 58.85 +/- 7.84 years. When CMBs were assessed as a continuous variable on T2*, the ICC3 reached 0.95 and 0.87 for the intra- and inter-rater assessments respectively. On SWI, the same parameters reached 0.87 and 0.65 for the intra- and inter-rater evaluations respectively (p always <0.001). Linear regression analysis between the absolute difference of counting and the mean of values, showed that the intra- and inter-rater agreements decreased significantly with the number of CMBs, especially when using the SWI sequence. For the intra-examination study, the R2 was 0.732, and the interaction between the mean value and the MRI sequence was significant with p = 0.032. For the inter-rater study, the R2 was 0.616, and the interaction between the mean value and the MRI sequence was significant with p < 0.0001. Results are detailled in Fig2. The absolute difference between CMBs counting against the mean count are displayed on Fig3. For the categorical classification of CMBs on T2*, the weighted-Kappa reached 0.97 and 0.74 for the intra- and inter-rater examinations respectively. For CMBs counting on SWI, these scores reached 0.94 and 0.83 for the intra- and inter-studies respectively. All tests were statistically significant (p<0.001). Because of the limited sample size, a category-based analysis could not be evaluated. The heatmaps presented in Fig4 showed a high intra-rater agreement. More discrepancies were found for the inter-rater, especially for T2* data.

The results of this study showed that, globally, the intra- and inter-observer agreements of the number of CMBs on T2* and on SWI were always relatively high, both for continuous and categorial countings. They also showed that the agreement for continuous countings decreased significantly with the number of lesions to be detected. As the SWI sequence is very sensitive to iron deposits, such an effect appeared larger for SWI than for T2*images. Because of the limited sample size, the potential effects of evaluation by number categories on inter or intra-rater agreement was not assessed.

Our results emphasized that the reliability of CMBs identification mainly depends on the burden of lesions to be detected. Future analysis will investigate whether an automated segmentation may improve the counting reliability [6]. Pending further in-depth studies, we suggest a cautious interpretation of results including high number of CMBs [8].

The placenta is a dynamic organ essential for supporting fetal growth and development, functioning as the critical interface for maternal-fetal exchange of nutrients, gases, and signaling molecules. Its normal development and function are crucial for a healthy pregnancy, yet can be compromised in conditions such as preeclampsia (PE) and fetal growth restriction (FGR), both of which are major contributors to perinatal morbidity and mortality (1). One commonly used preclinical model to study these conditions is the Reduced Uterine Perfusion Pressure (RUPP) model, which mimics key features of PE by surgically decreasing uterine blood flow (2). In this study, we use longitudinal Dynamic Contrast-Enhanced (DCE) MRI to non-invasively assess changes in placental perfusion across multiple gestational days in rats subjected to RUPP compared to normal pregnancies. This approach enables evaluation of both the acute effects immediately following the surgery and the potential compensatory adaptations that occur later in gestation.

Animal experiment: Our study used pregnant Sprague Dawley rats, divided into two groups: normal pregnancies (NP) (n=8) and RUPP pregnancies (n=10). The RUPP model was established by clipping the ovarian arteries and abdominal aorta to reduce uterine blood flow. Each animal underwent three MRI sessions starting at gestational day (GD) 14, immediately after surgery, followed by sessions on GD15 and on either GD16 or GD18. After the final MRI session, animals were sacrificed. 3D-DCE MRI: To assess perfusion, DCE MRI was conducted using a 7T scanner (MR Solutions, Guildford, UK), utilizing a 3D RF-spoiled gradient-echo sequence with the following parameters: TR/TE = 8/2 ms, flip angle = 25°, FOV = 60 x 60 x 60 mm³, matrix = 128 x 96 x 96. K-space filling was done using a pseudo-radial scheme in the two phase-encoding directions and a tiny golden-angle increment of 16.95 degrees between successive spokes. The pseudo-radial acquisition was repeated for 11 minutes. Subsequently, data was reconstructed with a temporal resolution of 22 seconds using a compressed sensing reconstruction algorithm with a total variation regularization in the temporal domain.

Semi-quantitative analysis of the DCE MRI signal intensity-time curves was performed using three perfusion parameters: area under the curve (AUC), peak enhancement, and time to peak (TTP). On GD14, placental perfusion was significantly impaired in the RUPP group compared to the NP group. Specifically, AUC was markedly reduced (RUPP: 52,000 ± 7,200 vs. NP: 70,500 ± 6,800; p < 0.001), peak enhancement values were lower (RUPP: 102 ± 15 vs. NP: 145 ± 10; p < 0.001), and TTP was significantly delayed (RUPP: 288 ± 21 ms vs. NP: 195 ± 18 ms; p < 0.01) At subsequent time points, perfusion metrics in the RUPP group showed progressive recovery. By GD18, AUC and peak enhancement values in RUPP placentas were comparable to those observed in the NP group (AUC: RUPP: 68,000 ± 8,200 vs. NP: 67,500 ± 7,400; Peak: RUPP: 140 ± 12 vs. NP: 138 ± 11). TTP differences also diminished by GD18.

The results demonstrate that placental perfusion was significantly reduced in the RUPP group compared to the normal pregnant group at the earliest post-surgical time point. This reduction was evident across all measured parameters, including AUC, peak enhancement, and TTP, indicating impaired placental blood flow following the surgical intervention. At later gestational days, the perfusion parameters in the RUPP group showed progressive changes. By the final time point, values for AUC and peak enhancement were similar between the RUPP and control groups, and the initial delay in TTP had largely resolved.

This study highlights the dynamic adaptive capacity of the placenta under RUPP procedure. Initial findings at GD14 showed significantly lower perfusion parameters in the RUPP compared to normal pregnancies. However, improvement by GD16 and GD18 suggests adaptive response that restore function over time. This result underscores the limited critical intervention window in the RUPP animal model and it provides insights into placental resilience and adaptation.

Neuro-Behçet’s syndrome (NBS) is the neurological form of Behçet’s disease. Machine learning (ML) and artificial intelligence (AI) are increasingly used in medical imaging to enhance diagnosis and prediction, particularly in radiology. MRI provides both traditional and advanced biomarkers—such as lesion characteristics and diffusion-based metrics—that help quantify brain changes. These biomarkers support predictive modeling even in rare conditions like NBS. Combining multimodal imaging with ML offers potential for personalized prognosis. Most research compares NBS patients to BS patients, with limited direct comparisons to healthy controls. A recent 2024 study [9] analyzed DTI data from 12 NBS patients and healthy controls, highlighting WM differences. This current study is notable for its relatively large NBS cohort from Türkiye, focusing on WM abnormalities using diffusion anisotropy indices and applying ML to identify key DTI metrics for accurately distinguishing pathological from healthy tissue.

Data were collected as a part of YOK 50006 "ML Based Differentiation of NBS with multi contrast MRI" project. project. This IRB approved study involved 14 NBS patients and 38 healthy controls, excluding one patient with a large frontal lesion. High-resolution T2-weighted MRI and diffusion tensor imaging (DTI) with 20 gradients were acquired. Three eigenvalue maps (λ₁, λ₂, λ₃) were generated. White matter (WM) masks were semi-automatically segmented from T2-weighted images using MIPAV’s level set method, starting at the axial slice where lateral ventricles appeared to reduce non-WM inclusion. T2W and B0 images were skull-stripped, bias-corrected, and co-registered via affine transformations using ANTS, with WM masks transformed accordingly and binarized. WM masks were semi-automatically segmented from T2-weighted brain MRIs using MIPAV’s level set method. Volumes of Interest (VOIs) were selected for both NBS and control groups, beginning segmentation at the axial slice where the lateral ventricles first appeared. This approach minimized non-WM inclusion, thereby reducing errors in diffusion tensor analysis. Voxel-wise WM diffusion indices—including λ₁, λ₂, λ₃, ADC, RD, FA, and RA—and their means were calculated. Min-max normalization was applied to histogram-based features (30 bins) of these variables to ensure comparability. These features served as inputs for classification. ML models were trained and tested using MATLAB’s toolboxes with leave-one-out cross-validation, evaluating features alone and in combination, with and without PCA (set to 5 components and 95% variance explained). ANOVA guided feature selection. Model performance was assessed via accuracy and ROC curves.

The study found that λ₁ and λ₂ was the most sensitive markers. ADC’s importance is driven by λ₁ and λ₂, while RD is influenced by λ₃, which is less distinct. FA and RA showed no significance, indicating that raw diffusivity metrics better detect WM abnormalities in NBS. Classification results demonstrated that λ₂, λ₃, and ADC achieved the highest accuracies, frequently exceeding 90%, with top performances reaching 98.08% for ADC (e.g., Fine Tree classifier). Quadratic SVM and Neural Networks also showed strong results, with accuracies up to 96.15%–100% after PCA application. PCA generally improved classification accuracy by 2–5%. In contrast, FA and RA consistently yielded lower accuracies, typically below 80%. These findings underscore λ₁, λ₂, λ₃, and ADC as the most sensitive and discriminative metrics for differentiating NBS patients from healthy controls. The 100% accuracy of the best model was shown in Figure 4.

The analysis revealed that traditional metrics (FA, RA) lacked significance for NBS, whereas raw eigenvalue features particularly ADC were highly effective. ADC alone reached up to 100% accuracy with Linear SVM and neural networks. Combining ADC with other metrics, especially using PCA, further enhanced performance. Ensemble methods, SVMs, and neural networks consistently showed strong results, with Linear SVM notably efficient. Despite some model instabilities, ADC proved crucial for accurate WM integrity classification and clinical application.

The analysis showed that traditional metrics like FA and RA were not significant for NBS, while raw eigenvalue features—especially ADC—were highly effective. ADC alone achieved up to 100% accuracy with models like Linear SVM and neural networks. Combining ADC with other metrics improved performance, especially with PCA. Ensemble methods, SVMs, and neural networks consistently performed well, with Linear SVM being notably efficient. Some models failed due to instability, but overall, ADC proved central for robust WM integrity classification and clinical potential. Statistical analysis reveals group differences, but lacks predictive power. ML enables individual-level classification by modeling multiple features, especially in high-dimensional medical imaging data like DTI.

Arterial Spin Labeling (ASL) perfusion MRI is a promising tool for detecting early biomarkers of neurodegenerative and psychiatric diseases [1]. However, significant physiological variability in cerebral blood flow (CBF) complicates its clinical use. Changes in mood may contribute to this variability and confound the interpretation of CBF measurements. This study evaluates the impact of mood on CBF using autobiographical recall with guided imagery during ASL scans.

ASL data in 48 healthy young adults (50% women, 18–29 years) were acquired during positive, negative, and neutral mood conditions, using a within-subject design (Figure 1). Mood was induced through guided imagery using autobiographical recall, trained before the MRI session. Perfusion was measured using a pseudo-continuous ASL with single-shot 3D-GRASE readout (labeling duration 1.5s, PLD 2s) on a single Siemens 3T Trio scanner at the UGent core facility GIfMI. ASL was processed using ExploreASL [2] and CBF was quantified for total gray matter (GM), anterior cingulate cortex (ACC), posterior cingulate cortex (PCC), hippocampus, anterior and posterior parahippocampal gyrus, amygdala, medial (pre)frontal cortex (mPFC) and putamen. Using linear mixed-effects models (LMM), the validity of the mood induction procedure, the GM CBF and spatial COV, and normalized regional CBF were tested for all regions. Multiple comparisons were corrected using the false discovery rate correction (FDR), and the significance level was set at alpha = 0.05.

Mood induction effectively altered self-reported happiness and sadness, with significant task-valence interactions for happiness (F(2,811)=166.38, p<0.001) and sadness (F(2,810.01)=217.67, p<0.001). A significantly higher total GM CBF for positive MIPs (mean(SE) = 0.8 (+-0.3) ml/100g/min) compared to negative MIPs (t(811)=2.62,p=0.025), but not neutral MIPs (t(811)=2.25,p=0.063) was detected. Spatial COV results showed no significant effects (F′s<2.31,p′s>0.099). For regional perfusion, no significant main effect of mood state was detected (F’s < 1.04, p’s > 0.247). Task effects (during mood induction versus during resting-state) were observed for PCC (F(1,811)=4.77,p=0.029) and mPFC (F(1,811)=5.18,p=0.023), though these became non-significant after correcting for multiple comparisons.

Mood changes were successfully induced during the MRI session. Positive mood induction significantly increased total GM perfusion compared to negative mood, while this effect was not detected for regional perfusion. However, compared to other perfusion modifiers such as age, exercise and sleep [3], these effects are relatively small.

In conclusion, a positive mood appears to significantly influence total GM perfusion. Albeit small effect size, mood state may influence cerebral blood flow and should be considered in future perfusion studies.

Pulmonary transit time (pTT) determined by cardiac magnetic resonance (CMR) has been shown to be impaired by various conditions such as heart failure [1], congenital heart disease [2] or surgical cardiac procedures [3]. Recently, pTT was also suggested as imaging predictor for cardiac involvement and prognosis in light-chain amyloidosis [4]. Frequently, pTT is determined from first-pass perfusion CMR as the time interval between the maxima of time-intensity curves in the right ventricle (RV) and left ventricle (LV). However, it has been suggested by Nelsson et al. [5] that the use of the center of gravity of the respective time-intensity curves might be more robust. It was the purpose of our study to compare the maximum-method and the center-of-gravity method with regard to discrimination between normal patients and patients with proven amyloidosis.

The study was approved by our local institutional review board. Patients (age > 18a, no contraindications for MRI) with proven amyloidosis and individuals without amyloidosis and no signs of cardiac impairment were included into this study. Subjects were investigated on a 1.5T MR scanner (MAGNETOM AvantoFit, Siemens, Germany). For pTT assessment, a first-pass perfusion sequence was applied, using a cardiac-gated single-shot 2D saturation recovery gradient echo sequence (true fast imaging with steady-state free precession - trueFISP); typical imaging parameters were as follows: repetition time/echo time: 2.2 ms/0.97 ms, saturation recovery time: 90 ms, flip angle: 50°, acquisition matrix: 128 × 82, field of view: 400 × 300 mm, bandwidth: 1370 Hz/pixel, slice thickness: 8 mm, parallel imaging mode: GRAPPA, and acceleration factor: 2. Image acquisition was started simultaneously with contrast media injection; a total of 60 images in one 4-chamber long-axis view as well as two short-axis views were acquired. Time-intensity curves were created with syngo.via software (VB80f, Siemens Healthineers) and circular regions of interest (ROI) were placed within the basal RV and LV blood pool on the first-pass perfusion 4-chamber long-axis view. Care was taken to avoid partial volume effects at the papillary muscles or myocardial wall. Average signal intensity within the ROI was plotted as a function of time, resulting in an indicator-dilution curve. Omitting recirculation, a gamma variate fit was applied to the data using the R Project for Statistical Computing 4.4.3 software (R Foundation for Statistical Computing). From the obtained fit parameters time to peak values (TTP) and time to center of gravity (CoG) values could easily be computed. pTT was then defined as the difference between the respective LV and RV values.

114 subjects were enrolled in this study, including 19 subjects, which were chosen as controls, without cardiac impairment and 89 patients with proven transthyretin amyloidosis (ATTR). Table 1 shows a summary of the obtained pTT values for heart-healthy subjects and ATTR amyloidosis patients for the different pTT methods. Wilcoxon rank sum test showed a significant difference between controls and patients with amyloidosis for pTT values of both methods (fig 1). ROC analysis resulted in similar pTT thresholds for both methods (9.55s for pTT(TTP) and 9.57s for pTT(CoG)). The area under the curve (AUC) for pTT(TTP) was 0.907 (95% CI: 0.849-0.965) with a specificity of 100% and a sensitivity of 71.9%. For pTT(CoG) the AUC was 0.668 (95% CI: 0.552 – 0.783) with a specificity of 84.2% and a sensitivity of 57.3%. As shown in figure 2, the difference between both ROC curves was highly significant (p < 0.001).

In this study pTT calculated by two different methods from first pass contrast enhanced CMR data were compared between controls and patients with proven ATTR amyloidosis. pTT(TTP) enabled a significantly better discrimination between the patient groups as compared to pTT(CoG). For pTT(CoG), variablity in contrast agent wash-out may lead to the observed decrease in discrimination between the patient groups.

pTT(TTP) obtained as difference between LV and RV TTP values were found to enable a significantly better discrimination between heart-healthy subjects and ATTR amyloidosis patients.

The His-Purkinje system is implicated in the initiation and maintenance of ventricular fibrillation, which can lead to sudden cardiac death [1]. This conduction network is essential for ventricular activation but remains poorly characterized and structurally variable across species [2]. Inhomogeneous Magnetization Transfer (ihMT) MRI [3–5], sensitive to myelinated structures [6], has recently shown promise for imaging the cardiac conduction system, particularly Purkinje fibers [7–9]. This study aims to optimize ihMT parameters for PF visualization, and complement it by histological validation.

Experimental setup: A left ventricular sample (Fig.1-a), containing myocardium and free-running PF was obtained from a female sheep’s (51.1 kg, 9 y.o) heart and fixed in 4% formaldehyde containing 0.1% of gadoterate meglumine (0.5 mmol/mL; Dotarem, Guerbet, France). For MRI, the sample was placed in a syringe filled with Fluorinert (Sigma-Aldrich; Fig.1-a), heated to 33±1 °C using water bath and continuously monitored with a temperature probe (SA Instruments, NY). 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 reception coil. A 2D ihMT-prepared RARE sequence (Fig.1-b) was used with the following readout parameters: TE/TR/PF/#M0/#MTw/TA=20ms/3000ms/1.8/10/200/2h and res=0.25×0.25×1 mm3. IhMT ratios (ihMTR%) were calculated as: ihMTR=100×(M_sing-M_dual)/M_0 where Msing=MT++MT- and Mdual=MT±+MT∓. MT+ and MT- refer to the MT-weighted images acquired with positive frequency offset saturation and negative offset saturation, respectively and MT± and MT∓ refer to MT-weighted images acquired with dual frequency offset saturation. IhMTR maps were generated using the script: https://github.com/lsoustelle/ihmt_proc. A range of different MT saturation parameters (Table 1) were tested to find the optimized contrast defined as: Contrast=ihMTR_PF - ihMTR_myocardium where a custom-made python script was used to draw two ROIs manually in PF (Fig. 1c, Mask1, red) and in myocardium (Fig. 1c, Mask1 green) to calculate ihMTR mean±SD. Further analysis was done by separating PF into free-running (Fig. 1c, Mask2, red) part of the fiber and the insertion point (Fig. 1.c, Mask2, blue), the part that enters the myocardium. Histological analysis: After MR experiments, histological analysis was conducted to provide detailed insights into the tissue structure. After dehydration, sample was embedded in paraffin and sectioned at 6 μm. Tissue sections were stained with Picro-Sirius Red for structural identification: collagen fibers, myocytes, and adipocytes were red, yellow and white, respectively. The polarized light (PL) filter (analyzer-polarizer, Nikon) allowed for differentiation between collagen type I (in red and yellow) and collagen type III (in green) [10].

Figure 2 illustrates ihMTR values in the PF (red) and myocardium (green) across different experiments. In all acquisitions, ihMTR was consistently higher in PF compared to myocardium. The protocols with the highest contrasts (namely, Pr1=2±1.2%, Pr2=2.1±1.3%, and Pr3=2±1.2%) are highlighted in the figure (dashed lines), and their corresponding MT saturation parameters are listed in the table. ihMTR values across the three protocols consistently showed higher signal in the free-running region of the PF vs. the insertion point. In Pr1, the values were 9.4 ± 1.6% vs. 6.6 ± 1.8%; in Pr2, 11.0 ± 1.9% vs. 8.1 ± 1.2%; and in Pr3, 9.4 ± 1.8% vs. 6.9 ± 1.6% for free-running vs. insertion point. For comparison, myocardium ihMTR values were 7.2 ± 0.7% in Pr1, 8.6 ± 0.8% in Pr2, and 7.2 ± 0.7% in Pr3. Fig. 3a shows histological views of the same sample under standard and polarized light (PL), along with the ihMTR map from Pr1 (Fig. 3b). The fiber contains Purkinje cells, collagen (red), and adipocytes (white), with adipocytes mainly at the insertion point and collagen more abundant in the free-running region. Under PL, Col I (red/yellow) is more prevalent than Col III (green); Col III is slightly more expressed at the insertion point.

Testing different ihMT protocols on one sample identified those with highest fiber signal and myocardium contrast. IhMTR was higher in the free-running PF region, which showed more collagen and Purkinje cells. The insertion point had lower ihMTR, likely due to adipocytes. IhMT effectively detects microstructural differences in the Purkinje network linked to tissue composition.

The Purkinje network is a variable and understudied structure. IhMT offers contrast sensitive to its heterogeneous makeup, influenced by collagen and adipocytes. This study establishes a framework for fiber imaging and validates it histologically. Future work will apply this approach to additional ex-vivo samples to better characterize ihMT contrast.

Glioma is the most common and aggressive primary cancer of the central nervous system[1][2]. Despite recent advances in clinical practice, the detection of prognostic biomarkers still requires invasive, complex, and costly methods[3][4]. Therefore, the need for a simple, reliable, and easy-to-use predictive model arises. MRI is a noninvasive modality known for its superior soft tissue contrast, making it particularly useful for diagnosing brain pathologies, including glioma. Radiomics is a field within medical imaging that specializes in the analysis and extraction of quantitative parameters from medical images. Given these insights, radiomics features extracted from MR images can unveil hidden patterns that cannot be identified on other imaging modalities, serving as essential building blocks for the development of prognostic prediction models and treatment decision tools in cancers. The goal of this study is to construct and validate radiomics prediction models based on preoperative gadolinium-enhanced T1- and T2-weighted images, and to evaluate their performance.

The UPENN-GBM (611 patients) and TCGA-GBM (135 patients) cohorts from TCIA (The Cancer Imaging Archive) were used in this study[5][6]. These datasets include magnetic resonance imaging (MRI) scans of de novo glioblastoma patients with gadolinium-enhanced T1-weighted, T1-weighted, T2-weighted, and FLAIR images. Automatic and manually corrected segmentations of tumor sub-regions—tumor core, enhancing tumor, and invasion—were available for all patients. UPENN-GBM was used for training, while TCGA-GBM served as an independent test set. Radiomics features were extracted from three tumor sub-regions (tumor core, enhancing tumor, and invasion), using gadolinium-enhanced T1 and T2-weighted MRI, processed with PyRadiomics v3.1.0[7]. For voxel-wise radiomics, K-means clustering was performed on the features with k = 3, aiming to segment the tumor into distinct imaging habitats. Habitat imaging refers to the identification of spatially heterogeneous subregions within a tumor that may reflect underlying differences in biological characteristics such as cellularity, vascularity, or hypoxia. These radiologically defined habitats provide a noninvasive means to capture intratumoral heterogeneity and are increasingly used to guide personalized treatment strategies[8]. The number of clusters was based on the three tumor sub-regions, and the area of each cluster was calculated in cm³. Pearson correlation analysis was applied to feature pairs, and one from each pair with r > 0.9 was removed to reduce multicollinearity. The remaining features, along with patient age, were used in a Cox proportional hazards regression model. All combinations of two and three predictors were tested. Four-fold cross-validation was used for robustness, and models were evaluated on the test set. Performance was assessed using P-values and the concordance index (C-index). Model simplicity was prioritized to avoid overfitting.

As representative example, the clustering result of the same radiomics feature for two different patients is shown in Figure 1. Thousands of models were built, trained, and tested based on the remaining features after applying Pearson correlation analysis. The C-index for the best models ranged from 0.682 to 0.689 (95% CI: 0.534–0.751; P < .0001) in conventional radiomics and from 0.665 to 0.667 (95% CI: 0.530–0.753; P < .0001) in spatial radiomics, with all models consistently based on three predictors. In conventional radiomics, the most predictive model was derived from features extracted from the invasion sub-region using T2-weighted images (Figure 2). Meanwhile, in spatial radiomics, the best predictive model utilized features from the whole tumor, extracted from gadolinium-enhanced T1-weighted images (Figure 3).

Both radiomics approaches offered comparable results, with the classical method having a slight edge over the spatial radiomics approach, and the same can be said for the two weighting mechanisms. However, there were differences in model performance when features were extracted from different tumor sub-regions, primarily due to tumor heterogeneity. Although these results could be further improved by including more predictors in the model, our goal was to keep the model as simple as possible, which will facilitate the reproducibility and generalization of the results. The next steps will be to improve model performance by implementing the elbow method to choose the optimal K-value for clustering, including more clinical information in the prediction model, and obtaining a second testing set from a local cancer treatment center in Lyon, France.

The implementation of radiomics in clinical practice can be helpful for survival prediction and risk stratification. Both classical and spatial radiomics can offer a simple, non-invasive and easy to use tool to help clinician in their decision making

Obesity is a complex, chronic condition influenced by sex and associated with alterations in brain microstructure, which can be non-invasively assessed using magnetic resonance imaging (MRI) [1]. While obesity has been linked to an increased risk of several cancer types [2], its specific role in the progression of brain tumors—particularly glioblastoma (GBM), the most aggressive primary brain tumor in adults—remains unclear. The aim of this study is to investigate MRI biomarkers that reflect the impact of high-fat diet (HFD) exposure at 10 and 20 weeks on GBM characteristics in a preclinical mouse model, while also assessing potential sex-dependent differences.

Adult C57BL/6 mice were randomly assigned to one of four groups based on diet type and duration: high-fat diet (HFD, 60% fat) or standard diet (SD), for either 10 weeks or 20 weeks (around 9 mice per sex, diet and diet duration group). Following the respective dietary period, GBM was induced in all animals via stereotactic injection of 10⁵ GL261 cells into the brain parenchyma. GBM development was followed-up by T2W MRI on a 7T system. Once tumor volume reached approximately 70 mm³, a multiparametric MRI protocol was conducted, including T2 and T2* mapping, magnetization transfer ratio (MTR), and diffusion tensor imaging (DTI). Parametric maps were generated with Resomapper, a custom-made Python toolkit, and four regions of interest: tumor core (TC), tumor periphery (TP), peritumoral zone (PZ), whole tumor (WT) and contralateral apparently healthy brain (HB) were manually segmented and quantified using ImageJ. Linear mixed effects models were used with R to test for the effect of area, diet, diet duration and sex and the interactions between them. Post-hoc tests were performed in case any of these were found, using FDR correction for multiple comparisons.

In all parametric maps, expected differences between regions of interest were observed. The HB area showed higher MTR and fractional anisotropy (FA) than the tumoral areas, and oppositely, higher values of T2, mean, radial and axial diffusivity (MD, RD, AD) were found in the tumor than in HB. T2* was more homogeneous, but slightly higher in the HB (Fig. 1). Regarding DTI parameters, in MD a significant effect of the interaction Area:Diet:Duration (p<0.05) was observed. No significant differences between diets or timepoints were directly found, but differences between areas vary depending on the diet group. Particularly, in the 10 weeks HFD group, regardless of sex, TP showed a significant difference with the TC that does not appear in other groups that even showed the opposite tendency (Fig. 2). The results in RD are similar to these. In AD only a Sex:Area effect was observed (p<0.05). Males tended to have higher AD values than females across all brain areas, specifically in the HB area, but differences were not statistically significant. Similarly, in FA some effects including sex were found but do not translate into significant group differences. T2 results showed many significant effects, including a strong effect of the Sex:Diet:Duration interaction (p<0.001) and also of Area:Diet:Weeks (p<0.05). At 10 weeks, there is a significant difference between diets in the WT area, as well as in the TP, but only in females. Specifically, tumor of HFD females had higher T2 values than the corresponding controls. At 20 weeks, this difference disappeared in the females but appeared in males. Also, a difference between diets in the HB area was observed only in males, being the opposite at 10 weeks than at 20 weeks: T2 values were higher in HB of the SD mice at the early stage and higher in HFD mice at 20 weeks of diet (Fig. 3). In T2* and MTR some significant effects including diet type and duration were found, but do not translate into significant differences between groups.

First of all, differences in all parameters between the HB and tumoral areas are the expected, indicating presence of edema and haemorrhage in all tumor areas, and also slight edema in the PZ [3]. The most interesting insights come from MD and T2. MD suggests that in the 10-weeks HFD mice the edema is more concentrated in the TP respect the TC, which does not happen in other groups. Moreover, the T2 results indicate that the evolution with obesity stage of the water content of the tumor happens differently between sexes, and in the case of males also of the HB area, which might be affected by the tumor differently. Further work is currently being performed to validate these results with immunochemistry markers of neuroinflammation and proliferation, as well as 1H HRMAS spectroscopy.

This work characterizes a murine model of GBM and obesity comorbidity, revealing that tumor characteristics and its effects on the rest of the brain are influenced by sex, diet type, and diet duration, despite numerous confounding factors. The immunochemistry and metabolomic spectroscopy assays will help understand the underlying mechanisms of these MRI results.

Brachytherapy (BT) for gynecological cancer treatment planning traditionally employs CT images for applicator and catheter reconstruction, and MRI imaging for organ and target contouring. Many centers now employ MRI-only BT planning. Reconstructing on MRI images poses challenges as applicators and catheters are more difficult to visualize on MRI than CT, and inherent MRI distortions can lead to dosimetric effects of 2-7% per mm of displacement. This research retrospectively assessed the dosimetric differences for gynecological brachytherapy treatment plans using applicators reconstructed on CT versus MRI images. Images from two MRI pulse sequences were evaluated: a T2-weighted-2D-PROPELLER and a T1-weighted-3D-LAVA-FLEX sequence.

The study included 12 cervical cancer patients undergoing three fractions each of BT with a Venezia (Elekta) applicator, either with or without catheters. Images from each implant were acquired on a 1.5 Tesla Avanto fit GE scanner and a Philips Big Bore Radiation Therapy CT scanner. The MRI scans consisted of a T2-2D-PROPELLER sequence and a T1-3D-LAVA-FLEX sequence. The MRI images were oriented in the plane of the tandem to minimize distortions.; thus, MIM (Medical Image Merge, version 7.2.8) was employed for reorientation. The Oncentra Treatment Planning System (Elekta, version 4.6.2) was used to reconstruct the applicators and catheters on each CT or MRI image. A 3D dose grid was calculated in Oncentra for each MRI or CT reconstruction using the dwell times and dwell positions from the clinical treatment plan. A gamma index analysis was employed to compare the dose maps between the reconstructions done on CT and MRI images, using a passing rate of 90% with a 3%/2mm dose difference/distance to agreement, and 10% maximum dose thresholding. Average gamma results were obtained for the T2- and T1-weighted images. A Wilcoxon signed ranked test was completed (significance level of 0.05, two-tailed) to assess if measured distortions using the T2-2D-PROPELLER versus the T1-3D-LAVA FLEX reconstruction are significantly different. Moreover, we investigated if the number of catheters affected gamma pass rates. A Wilcoxon rank-sum test was used to assess if one sequence outperformed the other for each number of catheters.

In some cases, the MRI sequences being investigated were not completed due to logistical issues. These cases were removed from the study. The T2- and T1-weighted images from two different studies were not included due to image corruption issues. In total, 27 T2-weighted images and 27 T1-weighted images were included. Figure 1 displays the reconstruction in example CT, T2-weighted, and T1-weighted images for a patient. For the Wilcoxon signed rank test comparing gamma results obtained from T2-weighted images to T1-weighted images, 26 total samples were included since the unusable images mentioned above were from different cases. Of these cases, 19 included catheters and the rest were applicator only. The average gamma result over all T2- and T1-weighted image reconstructions were 96.1 ± 2.8% and 96.6 ± 3.1%, respectively. Minimum gamma results were 90.2% and 90.5% for T2- and T1-weighted images, respectively. Figure 2 displays the skew of gamma pass rates for the T2- and T1-weighted images. Figure 3 reveals the gamma analysis results for T2-2D-PROPELLER and T1-3D-LAVA-Flex reconstructions versus the number of catheters. The Wilcoxon tests indicated there was no significant difference between T2- and T1-weighted reconstructions overall, or for 0 to 5 catheters. However, T1-3D-LAVA-FLEX significantly outperformed T2-2D-PROPELLER for cases with 6 catheters.

Figure 1 reveals that the T2-PROPELLER and T1-3D-LAVA-FLEX MRI images are of sufficient quality for reconstruction and contour delineation. All images passed the gamma index criterion of 90%. The results illustrate that there isn’t a significant difference between dose maps obtained from applicator reconstructions on T2-2D-PROPELLER and T1-3D-LAVA-FLEX MRI images compared to CT images. However, as seen in Figure 2, reconstructions done on T1-3D-LAVA-FLEX are skewed towards higher gamma rates compared to those done on T2-2D-PROPELLER. Notably, Figure 3 reveals that the mean T2-2D-PROPELLER minus T1-3D-LAVA-Flex gamma differential is within the standard deviation of the gamma pass rates (~ 3%) when 0 to 4 catheters are reconstructed. As catheter number increases to 5 or 6, T1-3D-LAVA-Flex reconstructions perform superiorly to T2-2D-PROPELLER reconstructions. T1-weighted images significantly outperform T2-weighted images for 6 catheters. These results suggest that T1-weighted images may be beneficial for image reconstruction, particularly if using more catheters.

T2-2D-PROPELLER and T1-3D-LAVA-FLEX MRI images are safe and effective for MRI-only gyne bracytherapy. T1-weighted images may offer benefits for reconstructions of 6 catheters or more.

Central nervous system tumors are the most common solid tumors in children, with about half arising in the posterior fossa [1,2]. Following surgery, approximately 25% of children with posterior fossa tumors develop cerebellar mutism syndrome (CMS) - a condition marked by delayed-onset speech, motor, and emotional disturbances typically 24-48 hours postoperatively [3,4]. Early identification of anatomical or functional changes, such as cerebral blood flow (CBF), is critical for understanding the pathophysiology of CMS. Intraoperative MRI provides a unique opportunity for monitoring of CBF during surgery. Arterial spin labeling (ASL) is a non-invasive MRI technique used to quantify CBF and is implemented during intraoperative MRI procedures at our center. However, CBF measurements can be influenced by various factors, including anesthesia, patient positioning, and MR acquisition parameters. To examine the methodological implications for interpreting intraoperative CBF values, we compared ASL scans acquired during surgery to those obtained the day prior in children with posterior fossa tumors.

Study design Six patients (2m/4f; age: 6.7±4.9 years) who underwent surgical resection for posterior fossa tumors were included. Written informed consent was obtained. Images were acquired using a 3-Tesla MRI scanner (Ingenia ElitionX, Philips Healthcare). Preoperative scans (timepoint 1) were obtained prior to surgery (2.8±2.7 days), using a 32-channel head receive coil with the patient positioned in supine position. Intraoperative scans (timepoint 2) were conducted in the operating room (prior to craniotomy), with patients positioned in surgical prone position, secured in a head clamp, and scanned using two single-loop RF coils to accommodate the intraoperative setting (Figure 1). ASL acquisition Pseudo-continuous ASL scan protocols were consistent across both timepoints, utilizing a 3D GRASE sequence with a labeling duration of 1800 ms and a post-labeling delay of 1800 ms. Imaging parameters were as follows: 8 dynamic scans; 4 background suppression pulses; matrix=64x64x23; FOV=240x240x161 mm3; acquired voxel size=3.75x3.75x7.0 mm3; TE=13ms; flip angle=90°. Repetition time (TR) differed between the two timepoints, which was 4280 ms at timepoint 1 and 4600 ms at timepoint 2. Image processing Quantitative CBF maps were generated using established recommendations [5]. The Bayesian Inference for Arterial Spin Labeling (BASIL) toolbox [6] was used to compute CBF values (mL/100 g/min). Inversion efficiency was set to 0.69 and T1 of blood was adjusted using patient-specific hematocrit values [7]. Adaptive spatial regularization and motion correction were applied, and voxel-wise calibration using M0 images was performed. To extract CBF values in cerebral gray matter (GM), tissue segmentation was performed on T1-weighted images using the FAST toolbox in FSL. GM maps were subsequently registered to native ASL space. Mean CBF values for cerebral GM were computed, and paired t-tests compared the mean cerebral GM CBF values between timepoint 1 and timepoint 2.

CBF maps for two patients are presented in Figure 2, illustrating systematic differences in perfusion. Quantitative analysis revealed that mean CBF values in the cerebral GM at timepoint 1 (50.16±26.11) were significantly higher than those at timepoint 2 (14.40±5.76) (p=0.01; Figure 3). Notably, this pattern of decreased CBF was consistent across patients.

In children with posterior fossa tumors, intraoperative ASL shows significantly lower CBF values compared to measurements obtained prior to surgery. This difference may arise from factors such as patient positioning (supine vs. prone). Studies in healthy adults have shown no significant CBF differences across these positions, except for a trend toward slightly lower CBF in prone position [8]. Similar comparisons have not been conducted in children. Intraoperative ASL imaging is limited by image acquisition constraints, including alignment along the anterior commissure–posterior commissure line, which may hinder the labeling slab from being perpendicular to the arteries, particularly when patients are in prone position. Additionally, due to time constraints, angiography is not feasible, further limiting accurate positioning of the labeling slab. These limitations reduce inversion efficiency, potentially leading to underestimated CBF values. These constraints complicate longitudinal CBF comparisons. Individual measurement of inversion efficiency, such as through phase-contrast velocity MRI [9], may help mitigate these effects. Alternatively, relative CBF maps can be used, although this will not preserve global CBF changes.

In children with posterior fossa tumors, intraoperative CBF values are significantly lower than preoperative values, potentially making comparison of absolute CBF measurements less reliable. Methodological factors, especially inversion efficiency, may contribute to the observed CBF differences.

Pancreatic cancer is one of the most aggressive cancers and tumors are characterized by hypoxia [1]. Hypoxia induces numerous adaptive cellular responses that contribute to tumour resistance to radiotherapy and certain drugs, and is a marker of aggressiveness [2]. To noninvasively map tumor hypoxia, several teams have explored MRI-based approaches, particularly Tissue Oxygen Level-Dependent (TOLD) contrast that uses oxygen as a contrast agent [3,4]. In this work, we attemp and validate by correlate on voxel wise with IHC marker CAIX. The aim of this study was to assess tumour hypoxia using TOLD contrast in mice subcutaneously grafted with pancreatic tumours, and to validate the method by voxel-wise correlation with reference histological imaging.

Two pancreatic cancer cell lines, AsPC1 and SW1990, were subcutaneously injected into the flanks of five mice per group. Once tumors reached 500 mm³, mice were anesthetized for MRI acquisition using a 9.4 T preclinical scanner. The center of each tumor was marked on the skin and overlaid with a water-filled tube to define the imaging plane (Fig. 1). The MRI protocol included a high-resolution T2w sequence, DW-MRI, and an inversion recovery fast spin echo multislice (IR-FSEMS) sequence. The IR-FSEMS sequence was acquired twice during air and twice during oxygen breathing for TOLD measurements. For TOLD analysis, R1 maps were generated by fitting IR-FSEMS data acquired under air (R1_air) and oxygen (R1_O2) conditions, and %∆R1 maps were computed as %∆R1=(mean(R1_O2) – mean(R1_air))/mean(R1_air). ADC maps were extracted from DW-MR images. Following MR scans, mice were euthanized, and tumors were excised and sectioned parallel to the imaging plane (Fig. 1). From each tumor, six 2 µm-thick histological sections were obtained at 150 µm intervals to capture the full histological information corresponding to the MRI slice thickness. Slides were stained with CAIX IHC, a marker of hypoxia, and a threshold-based segmentation approach was applied using QuPath to identify hypoxic pixels based on positive staining. The resulting CAIX-classified images were interpolated to the MR resolution, generating hypoxic fraction maps and then registered to the MR images using manual affine and automatic non-rigid registrations. Hypoxic pixels were identified on %∆R1 and CAIX fraction maps using the following thresholds: %∆R1 ≤ 0 and CAIX fraction >0.15.; all other pixels were classified as normoxic.

High ADC values were used to identify necrotic regions, which were excluded from the hypoxic pixel classification. Hypoxic segmentation maps derived from TOLD imaging were compared on a pixel-wise basis with those obtained from CAIX IHC. Evaluation maps were generated to assess the accuracy of TOLD-based classifications, distinguishing correctly identified normoxic and hypoxic pixels from misclassified ones. Figure 2 presents TOLD- (%∆R1) and CAIX- based segmentation maps alongside the corresponding evaluation maps for two tumors from each cell line (AsPC1 and SW1990). Figure 3a presents the percentage of pixels from the evaluation maps, indicating the proportion of correctly and incorrectly classified hypoxic and normoxic pixels based on TOLD imaging, across both cell lines and all tumours. Figure 3b displays the number of pixels identified as hypoxic or normoxic by the TOLD-based classification compared to the reference classification derived from CAIX-stained histological images.

In both MRI and histological assessments, the SW1990 cell line produced more homogeneous tumors, with a greater proportion of normoxic pixels, compared to AsPC1 tumors, which exhibited a higher prevalence of hypoxic areas. AsPc1 tumors displayed increased heterogeneity, with a more balanced distribution of hypoxic and normoxic pixels. This balance in AsPC1 tumors contributed to a higher classification error for TOLD imaging (51%) relative to SW1990 (29%), resulting in an overall error rate of 36%. Furthermore, the distribution of pixel counts across classification groups was consistent between TOLD and CAIX evaluations.

This study successfully established a workflow for validating an MRI-based method against histological reference images on a voxel-by-voxel basis, despite a substantial difference in resolution (0.46x0.46x2 µm3 vs 250x250x1000 µm3). TOLD contrast proved effective in assessing hypoxia in subcutaneously grafted pancreatic tumors in mice, showing a strong correlation with histological findings. This approach enables the evaluation of therapeutic efficacy in pancreatic cancer by measuring hypoxia, providing a valuable tool for future drug efficacy studies.

Pharmacokinetic (PK) models, a quantitative method of DCE-MRI to offer quantitative physiological markers of plasma volume fraction (vp), extravascular extracellular space fraction (ve), and permeability–surface area product (PS), have been employed to facilitate tumour stratification and treatment planning [1–3]. Extended Tofts Models (ETM), a category of PK models, is consisted of Fast Exchange Limit (FXL) and No Exchange Limit (NXL) approaches with proven utility in specific clinical scenarios, however the suitability for breast cancer imaging is yet to be established. We therefore set out to investigate the suitability of FXL and NXL approaches for breast cancer imaging using Monte Carlo simulation, and to explore the determinants for voxel wise confidence range.

We hence acquired DCE-MRI from a patient with invasive ductal carcinoma and conducted FXL and NXL analysis, and numerical simulation to derive voxel wise coefficient of variation (CV), with study design shown in Figure 1. The study was approved by the London Research Ethics Committee (Identifier: 17/LO/1777) and registered as a clinical trial [NCT03501394]. Imaging Experiment: DCE-MRI data were acquired on a 3T MRI scanner (Achieva TX, Philips Healthcare, Best, Netherlands), using a 3D T1-weighted spoiled gradient echo (SPGR) sequence, with a repetition time (TR) of 3.8 ms, echo time (TE) of 2.3 ms, flip angle of 12˚, voxel size of 1.0 × 1.0 × 1.5 mm³, and 29 dynamics. Image analysis was conducted using MRIcron (University of South Carolina, USA), with rigid-body motion correction applied. Maps of vp, PS, ve were computed separately for FXL and NXL analysis using SEPAL algorithm [4], with model T1 of 1.3 s [5], r₁ relaxivity of 5.9 s⁻¹·mM⁻¹, r₂ relaxivity of 17.5 s⁻¹·mM⁻¹ [6] and a population-averaged arterial input function (AIF) [7]. Numerical Simulation: The perfect signal for a voxel was generated using corresponding experimentally derived vp, PS, ve from FXL analysis. The baseline signal was computed as the average of the corresponding first 4 time points, and subsequently the time course of each voxel was normalised as the percentage difference of the baseline signal. The experimentally derived SNR of the same voxel from FXL analysis was employed to generate Gaussian-distributed noise and added to the perfect signal to simulate the realistic signal. The realistic signal was analysed using the FXL approach to yield simulation derived vp, PS, ve. The noise generation and fitting were repeated 100 times to create 100 sets of outputs and subsequently the mean, standard deviation and CV were computed. Maps of mean, standard deviation and CV for vp, PS, ve were obtained by iterating the simulation through all the voxels within the tumour. Subsequently, the maps were computed for NXL using experimentally derived vp, PS, ve and SNR and NXL analysis for fitting. Statistical Analysis: All statistical tests were conducted using SPSS (Release 29.0, SPSS Inc., Cincinnati, OH, USA). Wilcoxon signed-rank paired tests were performed to evaluate the differences in CV between FXL and NXL analysis. Spearman’s correlations tests were performed for CV against SNR and mean values.

There is no significant difference in vp, PS, ve between FXL and NXL analysis, and all the statistical findings can be found in Table 1.There was a significant difference in CV of vp (p < 0.001, Figure 3A) between FXL (0.0107 ± 0.1213) and NXL (0.0111 ± 0.1227). There was a significant difference in CV of PS (p < 0.001, Figure 3A) between FXL (0.0032 ± 0.1159) and NXL (0.0146 ± 0.2253). There was a significant difference in CV of ve (p < 0.001,Figure 3A) between FXL (0.0064 ± 0.0485) and NXL (0.0181 ± 0.1144). There was a significant negative correlation between CV of vp (p < 0.001), PS (p < 0.001), ve (p < 0.001) from FXL against SNR (Figure 3B). There was also a significant negative correlation between CV of vp (p < 0.001), PS (p < 0.001), ve (p < 0.001) from FXL against corresponding mean value (Table 1).

Although the mean PK model outputs showed no significant difference between FXL and NXL models with reasonable number of instances, the CV of FXL is significantly lower than NXL on all the output parameters, indicating FXL as the preferred model for measurement error consideration. Strong negative correlations between CV and both SNR and central parameter values indicate that signal quality and parameter magnitude jointly influence uncertainty. These findings highlight the importance of model choice and SNR-aware interpretation in voxel-level DCE-MRI analysis.

Voxel wise Monte Carlo simulations enable spatially resolved confidence range quantification of DCE-MRI in breast cancer. The FXL model showed lower variability compared to NXL model.

Distinguishing true glioblastoma progression/recurrence from treatment-related pseudoprogression/radionecrosis is a clinical challenge.[1] Conventional MRI routinely cannot differentiate between these entities, resulting in a risk of overtreatment or delay of necessary interventions.[2] Hypoxia targeted blood oxygen level-dependent (BOLD) MRI is a novel imaging technique that may allow for enhanced characterization of tumor tissue based on underlying vascular features and oxygenation metabolism. In this study we prospectively aim to test the hypothesis that the information provided by hypoxia-targeted BOLD MRI may help differentiate between tumor progression/recurrence and treatment effect changes.

Patients with radiographically suspected glioblastoma progression/recurrence were included to undergo a hypoxia targeted BOLD MRI. A computer-controlled gas blender (RespirAct) was used to induce transient standardized isocapnic hypoxia in the lungs. A T2* weighted gradient-echo (GE) echo-planar imaging (EPI) sequence was used to acquire the imaging under controlled oxygen modulation. A custom Matlab script and SPM were used for preprocessing and analysis of the images. The blood oxygenation level-dependent (BOLD) response of tumor regions was assessed and compared to standard contrast-enhanced (CE)-T1 and FLAIR sequences. The initial tumor board decision was extracted for each case of radiographically suspected progression. It was categorized into “progression/recurrence” and “pseudoprogression/treatment related changes”.

Six patients with radiographically suspected progression were included. Initial multidisciplinary discussion classified four as “progression” and two as “pseudoprogression”. In the progression group, contrast-enhancing regions showed consistent BOLD signal decrease in response to hypoxia. In pseudoprogression, larger areas of non-responsive tissue were observed. These preliminary findings suggest distinct BOLD signal patterns between the two entities.

This pilot analysis suggests that hypoxia-modulated BOLD MRI may differentiate progression from treatment-related changes based on differential response to hypoxic modulation during BOLD imaging underlying different hemodynamic and oxygenation features. While the sample is small, this approach builds on earlier CO₂-based BOLD imaging studies in radiation necrosis, though the underlying physiological mechanisms are different.[3] 18F-(fluoroethyl)-l-tyrosine positron emission tomography (FET-PET) is the current gold standard to detect progression.[4] Ongoing recruitment and planned correlation with FET-PET imaging and clinical follow-up will allow for more definitive assessment of diagnostic utility. The combination of hypoxia-targeted BOLD-MRI with PET and perfusion techniques bears potential to further complement radiological follow-up in treated glioblastoma patients. The included patients will be longitudinally followed to determine which lesions prove to be true progression.

Our initial observations indicate that hypoxia-targeted BOLD MRI may constitute an additional imaging biomarker for distinguishing true progression from pseudoprogression/radionecrosis in a multimodal imaging setting. Additional studies are necessary to evaluate BOLD responses in pseudoprogression cases. This strategy may enhance recurrence evaluation, resulting in more accurate future imaging of gliomas.

Accurate grading and identification of IDH status in astrocytomas are essential for guiding therapeutic strategies and predicting patient prognosis. Their 2021 WHO classification emphasizes the importance of molecular markers, particularly IDH status, in tumor characterization [1]. Considering this, further investigation is needed to explore non-invasive diagnostic approaches. On the other hand, since 2012 when specific sequences were developed to detect the 2-hydroxyglutarate metabolite [2,3] conventional MRS sequences are not regarded as an option to detect the IDH mutation in gliomas. This study aims to re-evaluate the potential of magnetic resonance spectroscopy (MRS) in determining IDH status (mt: IDH mutated, wt: IDH wild type) and tumor grade (2, 3, or 4) by analyzing metabolite patterns across different astrocytoma subtypes, under the assumption that the effects of mutation and grade on the metabolic profile will affect enough the whole spectral pattern to allow for non-invasive discrimination.

The study utilized Single Voxel (SV) MRS data collected at Hospital de Bellvitge with 1.5T Philips Ingenia and Intera scanners in Spain. Five classification tasks were performed on Short Echo (SE, 30 ms), Long Echo (LE, 136 ms), and concatenated SE+LE spectra. Sequential forward feature selection and linear discriminant analysis were used to identify features to distinguish between: 1/Astrocitoma-IDH-mt-grade-2 (A2-mt) from Astrocitoma-IDH-NEC-grade-2 (A2-wt), 2/Astrocitoma-IDH-mt-grade-3 (A3-mt) from Astrocitoma-IDH-NEC-grade-3 (A3-wt), 3/A2+A3-wt from A2+A3+ Astrocitoma-IDH-mt-grade-4 (A4-mt), 4/A2 from A3, and 5/ A2+3 from A4.Classifiers were evaluated with area under the curve (AUC) and Balanced Error Rate (BER). The best classifier was defined by the smallest BER in the training phase.

After application of inclusion criteria, there were 71 cases with SV MRS available for analysis at SE (15 A2-mt, 12 A2-wt, 20 A3-mt, 13 A3-wt and 11 A4-mt) and 69 cases at LE (14 A2-mt, 13 A2-wt, 18 A3-mt, 13 A3-wt and 11 A4-mt). Recurring discriminative spectral features include creatine, choline and lipids at SE and lactate and Glx, at LE which can be seen on the mean spectra (not shown). ). For the first question SE, and for the other questions the combined echo time, yielded the best classifiers. Notably, all AUC values for the best echo are quite high 0.938, 0.933, 0.834, 0.849 and 0.870 for the 5 questions, respectively.

Histologically verified low-grade gliomas can be identified non-invasively, prior to surgery, with a conventional magnetic resonance imaging exam, due to their anatomical and contrast uptake characteristics. However, it is not as easy to distinguish between grades or IDH status. For this reason, a biopsy is needed, being performed before or during the removal of the tumour, in order to characterise the particular type of glioma.The discovery of molecular biomarkers of favourable prognosis in glioma subtypes, such as the IDH mutation was taken into account in the 2021 revision of the WHO classification of brain tumors [4].

These findings support our claim that IDH status can be successfully determined for A2 and A3, while A2 can be differentiated from higher grades of astrocytoma without needing specialized sequences [5].

Brain metastases (BM) represent the most common malignant brain tumors in adults[1]. To obtain local control, neurosurgery and stereotactic radiosurgery (SRS) are competitive as well as complementary approaches that are further combined with systemic therapy[2]. Preoperative SRS offers the advantage of more precise target volume definition and reduces the risk of tumor spread into the cerebral spinal fluid space at the time of surgery[3]. Several MRI approaches have been taken to assess BM treatment efficacy, including synthetic MRI, quantitative MRI (qMRI), radiomics, magnetic resonance spectroscopy (MRS), and perfusion imaging[4–10]. With regard to SRS, treatment response is generally evaluated weeks after end of the treatment course[11]. For preoperative SRS, the availability of robust Quantitative Imaging Biomarkers (QIB) before surgery would be beneficial to assess treatment outcome and possible complications such as radiation necrosis on an individual patient basis. In this explorative study, we focus on early changes in conventional MRI contrast, which can be established from difference images calculated from pre-SRS and post-SRS MRI data. Image contrast in qualitative MRI-data (weighted-images) might be affected by hardware calibration, which may cause a bias in the difference images. This can be avoided using a quantitative MRI protocol (qMRI). Therefore, we used both, difference images from weighted MRI as well as qMRI to evaluate changes in BM following pre-operative SRS. In addition, we explored whether changes in automatic tumor segmentation mask volumes are consistent with the changes observed in the qMRI difference maps.

In vivo measurements were performed on six BM patients in two sessions (pre-SRS, post-SRS). The pre-SRS session was performed for SRS planning. The post-SRS session was performed within two weeks after SRS but prior to the surgical excision of symptomatic metastases. The measurements were carried out using a whole-body 3T MR Scanner (MAGNETOM Prisma (VE11C) or MAGNETOM Skyra (VE11C), Siemens Healthineers, Erlangen, Germany). In addition to the standardized Brain Tumor Imaging protocol (BTIP) [12], qMRI measurements were performed with a recently published vendor-based protocol[13]. Compared to the published protocol, the resolution was increased providing PD, T1, T2*, and QSM maps at 1mm isotropic. The planned radiation dose profile was later registered to the qMRI image space using FSL[14]. The BraTS-Toolkit was used to segment edema (OE), necrotic/cystic region (NE), and enhancing tumor (CE) masks. qMRI difference maps were computed by co-registering pre- and post-SRS qMRI maps, and calculating voxel-wise differences. For the regions WM, GM, NE, OE and CE, correlations were computed between the mean dose and the qMRI differences (post-SRS – pre-SRS). Fig. 1 shows an overview of the study and obtained MR data.

CE mask volumes increased for the majority of patients. With regard to OE mask volumes, 50% decreased and 50% increased, independent of glucocorticoid dosage between pre- and post-SRS MRI. These volume changes corresponded to the changes visible in the qMRI difference maps. qMRI changes did not significantly correlate with radiation dose. Figure 2 demonstrates differing changes in tumor segmentation volumes for two patients. As seen in Figure 3, changes in OE differ between patients, with some showing a decrease in edema and some showing an increase after SRS. Figure 4 demonstrates the group level box-plots of the mean qMRI difference values for the different tissue classes. Paired t-tests revealed no significant differences between WM+GM and NE, CE and OE classes.

While differences between normalized weighted-images revealed similar changes as for qMRI data for some subjects, it is not always possible to correct for all the transmitter and receiver field variations in the pre-RT and post-RT scans. qMRI data is preferred as QIB, as it ideally is independent of scanner- and session-related discrepancies. BRATs segmentations revealed variable changes in NE, CE and OE volumes across patients. These volume changes also correspond to the changes visible in the qMRI difference maps. The qMRI changes did not significantly correlate with radiation dose or with the cortisone-therapy dose, which implies that the observable change in the parameters is probably not just dose-dependent or cortisone-therapy dependent, but might signify individual tumor response to SRS. The different effects of SRS across patients may have predictive value for treatment outcome and possible complications such as radiation necrosis.

These results demonstrate that changes in qualitative and quantitative MR scans of BM patients are detectable even within two weeks after SRS.

Brain metastases (BM) are the most common malignant brain tumors in adults [1]. The therapeutic response to immune and targeted therapies remains unpredictable, however, reliable biomarkers could improve predictive accuracy. Recent advances highlight cerebrospinal fluid (CSF) profiling as a minimally invasive method to gain insights into the tumor microenvironment. Non-invasive metabolic imaging using MRS offers the potential to locally examine the tumor metabolism in vivo. In this study, edited MRS combined with qMRI is used to accurately measure absolute concentrations of lactate (Lac) not contaminated by lipids, glutathione (GSH) and gamma-aminobutyric acid (GABA) in BM. To our knowledge, for the first time the relationship between MRS-detectable Lac concentrations in BM and the Lac concentrations in CSF are analyzed.

Participants: 12 patients with BM: 4 with non-small-cell lung cancer, 3 with breast cancer, 2 with clear renal cell carcinoma, 2 with small-cell lung cancer and 1 with melanoma. Data were acquired using a 20-channel phased-array head coil on a 3T Siemens Prisma. Diagnostic imaging (3D T1w MPRAGE, 2D T2w, 3D FLAIR, 12 min) was performed, followed by MRS (15 min) and qMRI (8 min). Lumbar puncture was performed within 24 hours of MRI. MRS was acquired in the BM area (Fig. 1A) and in the contralateral normal appearing brain tissue (CL) (Fig. 1B). For MRS, MEGA-sLASER pulse sequence [2] was used with TR = 2 s and TE = 80 ms. The 90 Hz frequency selective editing pulses were applied at frequencies δLac/GSH = 4.56 ppm, δGABA = 1.9 ppm, δOFF = 7.5 ppm to acquire 3x64 ONGABA, ONGSH/Lac, and OFFAll acquisitions, respectively. 3 unsuppressed water spectra were also acquired. Voxels of 20x20x20 mm3 were used in 8 patients, while 25x25x25 mm3 voxels were used in 4 patients with larger BM. For the qMRI, a vendor-based protocol (Thomas et. al [3]) was used to obtain PD and T1 maps. T2 maps were estimated from the T2w: 1) T2w were corrected for receiver field inhomogeneities in SPM [4]; 2) from the corrected T2w and PD map, the PDw was synthesized (sPDw) considering the mean WM T2 = 80 ms [5]; 3) T2 map was computed: T2map = -TE/log(T2w_biascorr/sPDw) MRS preprocessing was performed in Gannet [6] (MATLAB). Fitting and water-referenced quantification were performed in LCModel [7] using appropriate basis sets. Absolute concentrations were quantified using the Gasparovich et. al. method [8], water relaxation correction was performed using voxel-wise water concentration, T1 and T2 values calculated from the PD, T1 and T2 maps, respectively, using the in-house MATLAB code. For metabolite relaxation correction, literature T1 and T2 values were used. Considering that the tNAA signal originates from normal brain tissue only, a PVfactor was calculated accounting for the partial volume (PV) of the BM to determine the BM’s Lac concentration. PVfactor = (tNAACL – tNAABM)/tNAACL

Examples of the spectra acquired in one patient are shown in figure 2. For the BM, mean (SD) of the Cr SNR and the linewidth (LW) were 18.9 (5.9) and 5.1 (1.4) Hz, respectively. For the CL, SNR was 25.2 (9.5) and LW was 5.3 (1.4) Hz. High consistency was observed between the CL spectra. The water suppression factor was always >99%. PV-corrected Lac concentration was 7.5 (8.2) mM. Figure 3A shows the fitting quality of the Lac signal (without macromolecules), which was significantly better in the BM compared to CL. No correlation was observed between the MRS Lac levels and the CSF Lac level (figure 3B). The GABA and GSH concentrations were lower in metastasis compared to the CL (figure 4).

In this study, MEGA-sLASER pulse sequence allowed to reduce lipid contamination of the Lac signal, providing high-quality spectra for Lac measurements. Furthermore, it allowed to get GABA, and GSH concentrations in BM, even with voxel sizes as small as 8 ml. The use of qMRI provided water relaxation parameters, which significantly impact metabolite quantification in lesions [9], including BM. The absence of a correlation between MRS-derived BM Lac and lumbar-puncture derived CSF Lac could be due to the different sources of these parameters. MRS lactate reflects localized intra- or peritumoral metabolism, indicating hypoxia, altered glycolysis, and inflammatory changes [10]. In contrast, CSF Lac may remain near-normal even in the presence of highly glycolytic parenchymal BM, especially if there is no leptomeningeal involvement. The reduced concentrations of GABA and GSH in the BM regions most probably indicate the loss of normal cerebral tissue. Data acquisition and processing are ongoing. To improve corrections for the PV effects by more accurately estimating the tissue type fractions within the voxel, BraTS [11] will be performed.

MEGA-sLASER combined with qMRI enables reliable Lac measurement in brain metastases. Voxel size limitations require partial volume effects correction. Yet, no correlation with CSF Lac was observed highlighting distinct metabolic sources.

The use of immunotherapy in patients with brain tumours has been challenging. The immunotherapy response assessment for neuro-oncology (iRANO) criteria [1] were proposed to address challenges of immunotherapy, but require multiple visits, and often fail to distinguish between progression, pseudoprogression and treatment related enhancement. This work aimed to evaluate serial quantitative MRI biomarkers [2] in patients with glioblastoma (GBM) who received immunotherapy in order to characterise early treatment response.

Eight GBM patients who received immunotherapy on a phase one clinical trial were assessed retrospectively. Each patient had an MRI examination at baseline and multiple post-treatment visits (range 2-6, depending on their evolution whilst on treatment). Cohort demographics are presented in Figure 1. MRI data were acquired between September 2019 and April 2024 using a clinical standard brain protocol including pre- and post-contrast T1w and diffusion-weighted imaging (DWI). Two scanners were employed in the study (1.5T Siemens, 6/8 patients for all of their visits; 3T Philips, 1/8 patients for all of their visits; one patient moved between scanners due to a system upgrade). Three volumes of interest (VOIs) were delineated on the POST-contrast T1w and DWI acquisitions (Figure 1): VOI 1 and VOI 2 delineated the tumoral volumes on T1w and Apparent Diffusion Coefficient (ADC) images, respectively; VOI 3 (drawn on ADC map) included tumour and any surrounding abnormal tissue (e.g. oedema). All delineation was performed in Osirix [3]. The T1w resolution (1x1x1 mm3) was resampled to DWI (1.8x1.8x5 mm3), to allow copying of VOI 1 drawn on POST-contrast T1w data to ADC map. Enhancement fraction (EF) maps were calculated: EF=(POST-PRE)/ (POST+PRE). Median values of EF and ADC were reported for each patient at each visit; volumes of each type of delineation were also recorded. Histograms were used to demonstrate serial changes of EF and ADC. Response was evaluated using iRANO criteria.

Figure 1 depicts the baseline images for all patients (ADC, PRE and POST T1w) demonstrating the tumour heterogeneity of the cohort. Two patients (4 and 7) started the treatment after surgery. A common characteristic of the remaining 6/8 patients is the extended oedema tissue surrounding the tumour (Figure 1). Best response for iRANO is shown for all patients. Overall, 5/8 patients demonstrated a rapid increase of tumoral volume (Figure 4) suggesting no response to treatment. These patients (3 to 7) stayed on treatment for shorter periods (58-201 days) than the responding patients (425-679 days). Large changes of serial ADCs of tumour (VOI2) were experienced by three patients (2, 7 and 8) (Figures 2 and 4 ). Patient 4 also experienced large changes of ADC (Figure 4), but such changes were considered less reliable due to the reduced size of the tumour. Patient 2 (orange line in Figure 4, CR by iRANO) showed a tumour ADC increase for several post-treatment visits (Figure 3) corroborated with stable tumoral volumes (Figure 4) suggesting a good response. At the same timepoints, EF has the largest increase suggesting non-response, which, based on the ADC changes, may be considered pseudo-progression. Until visit 6, patient 8 (grey line in Figure 4, SD by iRANO) demonstrated increased ADC and stable EF and tumour volume, suggesting continuous good response. At visit 6, a large increase in tumour volume and an ADC decrease were observed, indicating non response. This change of treatment response is weakly suggested by the EF biomarker (showing a very small increase). Patient 7 (pink line, PD by iRANO) demonstrated an example of a non-responder with large decrease in ADC and small increase in EF after treatment. (Figures 3 and 4). Over the whole cohort, the serial ADCs of VOI3 (tumour+oedema) experienced smaller change than that of ADCs of VOI2 (tumour).

This descriptive study is limited by the small number of patients, the tumour heterogeneity and the use of two scanners during the trial. Despite these limitations, the ADC biomarker was a better predictor of treatment response than EF in both responders and non-responders. The ADC biomarker was able to identify pseudoprogression for the first two post-treatment visits of patient 2, which was confirmed histologically. Moreover, the early increase in ADC observed for the best responder (patient 2) might suggest ADC as a promising biomarker for early treatment response assessment. This study highlights the need to include quantitative MRI markers in treatment trials in order to prospectively evaluate which are the most informative.

In this descriptive study, ADC biomarker showed promising results in characterising treatment response for patients with GBM above tumour volume and enhancement fraction alone and should be characterised further in a larger cohort of immunotherapy clinical trials.

Glioblastoma (GBM) is the most common and aggressive primary brain tumor in adults, with a survival rate of 9–18 months [1]. Following treatment, patients may present with either true tumor progression (TP) or pseudoprogression (PsP), a treatment-related effect. Differentiating between TP and PsP is critical for appropriate clinical decision-making[1–3]. While conventional magnetic resonance imaging (MRI) plays a crucial role in brain tumor assessment, it is limited in distinguishing TP from PsP based solely on morphological imaging. Advanced techniques, such as diffusion-weighted imaging (DWI), offer additional information that may aid in this differentiation[2, 4, 5].

This retrospective study evaluated 49 patients diagnosed with GBM who underwent therapy and follow-up MRI scans. On apparent diffusion coefficient (ADC) maps, three regions of interest (ROIs) were drawn on the largest lesion area visible on contrast-enhanced T1-weighted images, and three ROIs were placed in the contralateral normal-appearing white matter. From each ROI, ADC mean (ADCmean), maximum (ADCmax), and minimum (ADCmin) values were extracted. Subsequently, ratios of ADC values between lesion and normal tissue (rADCmean, rADCmax, rADCmin) were calculated.

Patients with TP showed lower ADC values compared to those with PsP. ADCmean, ADCmin, ADCmax, rADCmean, and rADCmax values were significantly different between the TP and PsP groups, while rADCmin did not show significant differences. White matter ADC values were significantly lower than lesion values but did not differ between TP and PsP groups. No significant differences were observed among the three lesion ROIs for ADC metrics within each group. In the white matter, all ADC metrics for the TP group and ADCmax and ADCmean for the PsP group were consistent across ROIs, except for ADCmin in PsP patients. ADCmax (AUC = 88.9%) and rADCmax emerged as the best metrics for distinguishing TP from PsP, with respective cut-off values of 1.822 × 10⁻³ mm²/s and 2.141. Both metrics achieved 83.3% sensitivity, 100% specificity, and 85.4% accuracy.

Our findings highlight that DWI-derived ADC metrics, particularly ADCmax and rADCmax, are valuable biomarkers for differentiating TP from PsP. The significant differences in ADC values reflect underlying pathophysiological variations between true tumor growth and treatment-related changes[2, 4–6]. The high specificity and diagnostic accuracy demonstrated by ADCmax and rADCmax underscore the clinical potential of incorporating DWI analysis into routine GBM follow-up imaging protocols.

DWI provides significant added value in the post-therapy differentiation between TP and PsP in GBM patients. ADC metrics, especially ADCmax and rADCmax, show strong diagnostic performance and can support more accurate clinical decision-making.

Ultra-high-field magnetic resonance imaging (MRI) provides enhanced spatial resolution and tissue contrast, offering new opportunities for detailed tumor characterization [1]. However, current clinical MRI techniques often fail to reflect the heterogeneous nature of tumor tissue, limiting microstructural assessment [2]. Frequency-dependent multidimensional diffusion-relaxation correlation MRI (ωMDR-MRI) offers a novel framework to resolve sub-voxel features and characterize complex cellular environments. Recent nonparametric metrics further enable consistent assessment of tumor regions across scans and field strengths [3]. This study utilizes ωMDR-MRI at high resolution to improve detection of tissue anisotropy and microstructural features in gliomas, aiming to enhance both surgical precision and broader treatment planning.

A 74-year-old male glioma patient (WHO grade 4, IDH wildtype) was scanned preoperatively with a 1.5T Siemens MRI scanner. Whole tumor, contrast-enhancing tumor, and necrotic regions were segmented using 3DSlicer [4] from pre- and post-contrast T1-weighted MRI (TR = 2000 ms, TE = 2.69 ms, voxel matrix = 256 × 256 × 176, voxel size = 0.9766 × 0.9766 × 1 mm³), as shown in Figure 1A. Segmented voxels were z-score normalized, and histogram properties were analyzed for comparison. For visualization, the patient received Gliolan preoperatively. Following resection (Figure 1B), ex vivo scans were acquired at 7T and 11.7T using Bruker MRI systems, with parameters in Table 1. Among the nonparametric maps derivable from ωMDR-MRI, squared normalized anisotropy was used due to its normalized formulation, enabling robust comparison across field strengths. It was calculated as the squared difference between parallel and perpendicular diffusivities, normalized by their sum, quantifying microscopic diffusion anisotropy from the frequency-dependent diffusion tensor. Five anatomically relevant ROIs were manually defined on ex vivo T1-weighted images: supramarginal zone of white matter (red), supramarginal zone of gray matter (green), tumor border (blue), and tumor regions (purple and yellow). Each ROI was co-registered via affine transformation (ANTs [5]) onto the anisotropy maps from both 7T and 11.7T (Figure 1C). Voxel-level comparisons between field strengths were performed using Wilcoxon signed-rank test with Dunn’s posthoc test. Within-field ROI comparisons were made using the Friedman test followed by pairwise Wilcoxon tests. Kernel density estimation (KDE) visualized signal distributions to highlight differences in central tendency and shape.

Signal intensities from preoperative T1-weighted MRI showed statistical differences across tumor regions (Table 2A). Enhancing tumor areas had the highest mean intensities, followed by whole tumor and necrosis. The whole tumor showed the highest skewness, suggesting heterogeneity, while enhancing and necrotic areas showed lower skewness, indicating more balanced distributions. Pixel intensities in white matter, gray matter, and tumor regions differed significantly between 7T and 11.7T (p<0.01) and within each field strength (p<0.01) (Table 2B). No significant difference was observed between the tumor border and tumor regions at 7T; however, at 11.7T, the border region had significantly lower anisotropy (p<0.01). Signal intensities generally decreased at 11.7T (Figure 2). In white matter, distributions became more symmetric and less flat, suggesting improved homogeneity. Gray matter showed increased skewness and kurtosis, indicating higher variability. The tumor border showed flatter, more asymmetric distributions, suggesting structural heterogeneity. In tumor regions, increased positive skewness at 11.7T pointed to better detection of intratumoral complexity.

We found increased homogeneity in white matter, higher signal variability in gray matter, and clearer structural heterogeneity in the tumor border region at 11.7T. These results highlight the improved sensitivity of ultra-high-field imaging to microstructural features. The tumor region’s positively skewed distribution supports better detection of intratumoral complexity at higher field strengths. Additionally, while 7T did not differentiate between the border and tumor regions in terms of anisotropy, 11.7T revealed significantly lower anisotropy in the border zone, suggesting increased sensitivity to peritumoral heterogeneity. These findings emphasize the potential of ultra-high-field imaging to enhance tissue characterization in challenging regions. Future steps of this multiscale study will include microscopy image integration to assess cellular and subcellular structures, offering a refined understanding of the tumor microenvironments.

This study demonstrates that ultra-high-field MRI at 11.7T provides improved sensitivity to tumor tissue microstructure and may support more accurate diagnosis and treatment planning in gliomas.

Soft-tissue sarcomas (STS) are rare tumors with significant inter- and intra-tumor heterogeneity. Consequently, there is a need for imaging methods to guide biopsy[1], targeting heterogeneous components, and for evaluation of novel therapies. MRI enables quantitative assessment of tumors but better understanding of the link between quantitative MRI and tumor microstructure is required. Developments in image-guided biopsy[2] provide an opportunity to target biopsy sites using robotic guidance combined with quantitative MRI, while advances in digital pathology including multiplex immune-fluorescence (mIF) enable rapid characterization and quantification of cells in biopsy samples[3]; combining these techniques provides an opportunity to evaluate quantitative MRI parameters and corresponding tumor microstructure. The aims of this study are (a) develop a workflow combining image-guided biopsy with digital pathology in STS; (b) use this methodology to evaluate correlation between apparent diffusion coefficient (ADC) and numbers of cells, and between enhancement fraction (EF) and endothelial cells.

Patients were recruited at one centre as part of a prospective study[2] with written consent for use of tissue samples. Patients underwent an MRI examination before biopsy (1.5T MAGNETOM Sola, Siemens Healthineers, Forchheim, Germany) described in Fig.1. Antibodies were optimized for immunohistochemistry (IHC), multiplex positions and validated with single-plex IF (Opal TSA system, Akoya). Controls were used for IHC and IF optimizations. Two panels of mIF were optimized. Panel A was used to identify tumor cells, T-cells and B-cells. Panel B was used to identify macrophages, endothelial cells and cancer-associated fibroblasts (CAFs). Panels are described in Fig.2. Whole-slide images were generated using a Leica Bond autostainer and mIF images were acquired using a PhenoImagerHT (Akoya). Image analysis was done using HALO image software (IndincaLabs). Panels A and B were analysed for each biopsy sample to output the total number of cells and number of tumor cells, tumor-infiltrating lymphocytes (TILs=T-cells+B-cells), macrophages, endothelial cells and CAFs per unit area. Correlation between cell types was assessed using Pearson correlation coefficient. Correlation between ADC or EF and number of cells per unit area was assessed initially using Pearson correlation coefficients across all samples. Further analysis used a linear mixed effects model to evaluate the effect of each cell type on ADC, with number of cells per unit area as fixed effects (after log transformation and normalization) and patients as a random effect (fitlme, Matlab2021a). The variance partition coefficient (VPC) was used to assess inter-patient variation in ADC[4].

32 biopsy samples from 11 patients (2-3 samples per patient) were included. Median ADC across all samples was 1570x10-6mm2/s (range (656-2830)x10-6mm2/s). Median/range of each cell type are shown in Fig.3. Correlation between number of each cell type per unit area was weak-moderate (r=0.2-0.6) except between macrophages and CAFs, which were strongly correlated (r=0.7). ADC showed negative correlation with total number of cells per unit area and negative correlation with number of tumor cells, macrophages and CAFs per unit area; correlation between ADC and TILs or endothelial cells was weak (Fig.3). Mixed effects modelling included tumor cells, macrophages and CAFs (excluding TILs and endothelial cells). Number of CAFs per unit area had a significant effect on ADC (p=0.01) but numbers of macrophages and tumor cells did not have significant effects (p=0.7, p=0.9). VPC was <0.1. No correlation was observed between EF and number of endothelial cells per unit area (Fig.4).

This study demonstrated a workflow combining quantitative MRI, image-guided biopsy and digital pathology for validation of imaging biomarkers and investigation of tumor heterogeneity. Correlation between ADC and total number of cells per unit area confirms that ADC is inversely related to cellularity in STS. However, the significant effect of CAFs on ADC in the mixed effects model shows that interpretation of ADC estimates depends on tumor stroma, not solely on tumor cells. The small VPC suggests inter-patient variation is small compared with total variation across all samples. The absence of correlation between EF and endothelial cells is consistent with previous studies[5] suggesting that enhancement depends on the function of the tumor vasculature rather than presence of vessels. A limitation of the study was the small sample size. It was not possible to analyse subtypes separately as 9/11 tumors were liposarcomas; future studies will investigate other subtypes.

ADC is correlated with total number of cells per unit area but the number of CAFs per unit area is also significant indicating the importance of stroma in STS. This study demonstrates a workflow combining quantitative MRI and image-guided biopsy with digital pathology in STS.

Mini-invasive MR-guided Laser Interstitial Thermal Therapy (LITT) is used to treat tumors in various organs (brain [1], liver [2], prostate [3]) using one or more optical fibers [4–6]. However, clinical procedures often rely on predetermined laser power and exposure time, which can lead to the target area being over- or under-heated. A multi-probe laser device is presented, whose output power is automatically regulated in real-time from MR-temperature maps to force tissue temperature to follow a predefined desired temperature-time profile.

LITT device: A prototype multichannel LITT system (Alphanov, France) was connected to three independent optical fibers (400 µm diameter), each terminated by a diffuser tip (2 cm length, 1.8 mm diameter). Fibers were inserted into the leg muscle of an anaesthetized pig (~35 kg, ethics approved), forming a triangle with ~5 mm spacing. Each fiber was independently connected to a 976-nm laser diode that was controllable and interfaced with the thermometry pipeline [7]. MR-Imaging workflow: The procedure was performed on a 1.5 T clinical scanner (Avanto, Siemens, Germany) using a 4-channel surface coil placed above the leg and two elements embedded into the MRI bed coil) for signal reception. The acquisition protocol included: (1) anatomical imaging for probe positioning, (2) real-time MR thermometry during treatment, and (3) post-ablation assessment. Anatomical scans were acquired using a 3D MPRAGE (TI=1100 ms, TE=3.3 ms, TR=2000 ms, FA=15°, voxel size: 1.17×1.17×1 mm, FOV: 300×253×120 mm), before and after probe insertion into the muscle. A stack of 10 slices was positioned over the probe tips to perform PRFS-based MR-thermometry [8]. Phase images were acquired at 1 Hz using a single-shot EPI [9] (TE=21 ms, TR=1000 ms, FA=50°, voxel size: 1.56×1.56×3 mm, GRAPPA=2, partial Fourier=6/8, bandwidth=1150 Hz, FOV: 200×200 mm). Thermal maps were computed in real time and displayed on the fly ( Certis Therapeutics software). After completion of the thermal treatment, non-perfused volumes were visualized from 3D T1-weighted images (TE=2.16 ms, TR=4.49 ms, FA=10°, voxel size: 1.19×1.19×1.2 mm, FOV: 308×380×120 mm) acquired one minute after intravenous injection of gadoteric acid injection (0.5 mmol/kg). Regulation algorithm: The regulation algorithm was implemented in Gadgetron to automatically control heat deposition in real-time in several Regions of Interest (ROIs) centered on each laser. Input parameters: (1) the temperature-time profiles defined for (2) each ROIs (3x3x10 pixels each) positioned around each diffuser, (3) tissue thermal parameters (absorption and thermal diffusivity) determined from a calibration step, and (4) the heating pattern of each laser source. A PID controller combined with the Bio-Heat Transfer Equation was implemented [10]. At each acquired temperature stack [11], the power of each source was updated. Calibration step: Before temperature regulation, a constant power of 6 W was sequentially applied under MR-thermometry monitoring during 30 s on each diode, interleaved with a cooling period of 50 s, and the hottest voxel in temperature maps associated to each fiber was automatically detected and served as input for automatic ROI positioning. The temperature fitting method based on the BHTE equation was used to estimate the absorption and thermal diffusivity (D) [10], as well as the source function Si [11].

Figure 1 shows an example of 3D scan performed after inserting the laser probes into the muscle. The stack of slices for MR-thermometry is indicated by blue rectangles. Figure 2 shows examples of automatic temperature regulation with (Top) an identical target heating profile applied to the 3 ROIs (+30°C, 275 s); and (Bottom) different heating profiles applied on 2 ROis: a single plateau (+25°C, 180 s) applied on the first ROI, and 2 plateaus of +15°C and +30°C on the second ROI, both lasting 60 s. Table 1 summarizes metrics to evaluate the regulation algorithm performances. For both experiments and all ROIs, the mean ± std of the difference between target and measured temperatures remained below 0.3 ± 1.7 °C. Figure 3 shows the temperature and thermal dose images for all slices, illustrating the possibility to modulate the size and shape of the treated volume.

The MR-thermometry based automatic regulation is precise enough to force the temperature to follow predefined profiles during multi-probe LITT.

This multi-region control of heat deposition enables conformable therapy and may prevent overheating of critical structures in order to increase MRI-controlled treatment efficiency and safety.