Magnetic resonance imaging (MRI) techniques that exploit dipolar order-related contrasts, such as inhomogeneous magnetization transfer (ihMT) [1,2], show promise for monitoring de-/remyelination processes [3,4]. A pivotal factor influencing the ihMT contrast is the difference in dipolar order relaxation time(s) (T1D) between tissue’s constituents. In particular, myelin’s ‘long’ T1D inherits a promising application potential, however, the lack of understanding of dipolar relaxation in biologically relevant systems prevents thorough interpretation of the resulting contrasts. The impact of membrane dynamics [5], i.e., the ensemble of motional processes occurring at different spatio–temporal scales, on T1D-relaxation remains largely unknown. Myelin membranes are primarily composed of lipids in a lamellar liquid-ordered state [6] characterized by a well-known motional landscape [7,8]. Using a ‘refocused’ Jeener–Broekaert (JB) sequence, we analyse the temperature-dependence of T1D-relaxation by detecting the non-aqueous proton signal in liposomes that mimic myelin lipid membranes [9-11]. The aims are two-fold: To decipher the membrane motional processes driven T1D-relaxation and to elucidate the influence of prominent membrane constituents such as phospholipids (POPC), cholesterol (CHOL), cerebrosides (CER) and Myelin Basic Protein (MBP) on T1D-relaxation.

Sample preparation: Multi-lamellar vesicles (MLV) (Fig. 1) were prepared according to an established procedure [12,13]. (1) Lipids (POPC, POPC-CHOL (molar ratio of 0.6/0.4), POPC-CHOL-CER (0.4/0.4/0.2) / + MBP) were dissolved in CHCl3 (evaporated, water added, freeze-dried); (2) resulting powders were hydrated with D2O (hydration ~80% (w/w)) and underwent three freeze-thaw cycles; (3) Lipid dispersions were transferred into a 4-mm MAS-NMR rotor. NMR measurements: Acquisitions were performed on a solid-state Bruker Avance III 500 MHz spectrometer. To quantify T1D(s), a chemical-shift refocused JB sequence (Fig. 2) was conducted at different temperatures (273 K to 315 K) [11]. Briefly, the sequence comprises a pair of phase-shifted RF pulses that partly transfer Zeeman into dipolar order. The unwanted off-resonance signal contributions from ‘mobile spins’ are reduced by a 180° RF pulse (Fig. 2). Analysis: Raw data files were pre-processed using MATLAB routines (Fig. 2). JB signal decays were obtained by integrating the positive lobe of the ‘anti-phase like’ spectra measured at different evolution times τ. Two decay functions were utilized to model the non-monoexponential JB signal decays: A sum of n=3 exponentials and a sum of two modified stretched exponential functions (Fig. 2). Evolution times τ shorter than τ<200 μs were not considered, since a time τ~T2 is required to establish a novel internal equilibrium state. An ‘effective’ dipolar order decay time was calculated and used for the activation energy analysis of R1D by a simple physical model (encompassing only two distinct statistically independent motional contributions and assuming that each type of motion can be characterized by an ‘effective correlation time’ τc, details in Fig. 4)

MLVs of pure POPC exhibit a different temperature trend compared to the other lipid model membranes (Fig. 3). While POPC shows a maximum of the effective of ≈23 ms in the vicinity of 295K, the cholesterol containing liposomes (‘liquid-ordered phase’, LO) exhibit a roughly linear trend of as a function of temperature. In the LO-phase, an increase of for higher temperatures could be observed. Consequently, the longest T1Ds are obtained for the highest temperature (315 K) with values of about 30 ms for POPC-CHOL, less than 10 ms for POPC-CHOL-CER and about 25 ms for POPC-CHOL-CER+MBP.

The analysis yielded two markedly different activation energies and correlation times (Fig. 4): Motion ‘1’ with lower a Ea^1 (~15–25 kJ/mol) and longer τc^1 (10-5 to 10-3 s), and motion ‘2’ with higher Ea^2 (of 40–80 kJ/mol) and shorter τc^2 (10-9 to 10-7 s). According to the motional landscape for lipid model membranes [7,8], motion ‘1’ can be assigned to collective motions, i.e., membrane undulations (hydrodynamic µm-scale deformation) of the entire lipid membrane, whereas motion ‘2’ reflects intermolecular motions, i.e., motion of the lipid molecules as a whole (nm-scale). Within the biologically relevant temperature range (20–40°C), slow collective motions are the primary contributor to T1D-relaxation in all myelin model systems.

Our results indicate that dipolar relaxation is highly sensitive to the chemical composition of myelin membrane models. Slow collective motions are the main drivers of T1D-relaxation (within 20–40°C). The presence of MBP strongly attenuates membrane undulations, thereby slowing down T1D-relaxation. This opens up new perspectives for characterising alterations in myelin-related neurodegenerative diseases and should lead to major advances in using T1D-weighted contrast for myelinated tissues in MRI.

Magnetic field correlation imaging (MFCi) quantifies microscopic magnetic field inhomogeneities arising from magnetic susceptibility variations in tissue [1,2,3]. MFC is defined as the variance of local B0 perturbations, and can be expressed through the autocorrelation (e.q 1), where B(t) represents the local magnetic field experienced by diffusing spins at time t [4]. In white matter (WM) myelinated axons induce an anisotropic local frequency shift due to the radially oriented lipid bilayers around the axis of axons. This frequency shift scales as sin²(θ), where θ is the angle between the fibre axis and the direction of B₀ [5,6]. Since the MFC is proportional to the variance of local field perturbations, we predict that the MFC will scale with sin⁴(θ). Demonstrating this orientation dependence would provide strong first evidence that the MFC in WM is driven by the myelin-water microenvironment rather than other sources, such as iron, and promote MFC as a new imaging marker of myelin status. The aim of this study is to use ultra-high field (UHF) (≥7 T) to test this prediction in-vivo in humans across 20 canonical WM tracts.

Ten healthy controls were scanned on a Siemens Magnetom 7 T+ with a 32-channel headcoil using an EPI asymmetric spin-echo sequence (TE = 46 ms; refocusing pulse offsets tₛ = [0, 3, 6, 9, 12] ms; TR = 3000 ms). MFC maps were derived by fitting the noise-aware nonlinear signal model (eq. 2), where σ is the standard deviation of residual noise following preprocessing. Individual MFC maps were nonlinearly registered to MNI152 space and averaged into a group mean volume to maximize signal-to-noise ratio (SNR). WM fibre orientation was characterized using the FSL HCP1065 population-averaged principal eigenvector (V₁) atlas [7], from which θ was derived voxelwise through arccos(|V₁ ⋅ B₀|). Twenty WM tracts were defined using the JHU ICBM tract atlas (25% probability threshold) [8]. Within each tract, sin⁴(θ) values were binned (N = 100) and Spearman rank correlations were computed between the binned mean sin⁴(θ) and binned mean MFC (Figure 1).

Of the 20 WM tracts investigated, 14 (70%) showed a statistically significant positive Spearman correlation between the binned mean MFC and binned mean sin⁴(θ) (Figure 2). The median ρ across all tracts was +0.672, and the voxel-weighted mean ρ was +0.574. A sign test confirmed a significant directional bias toward positive trends (p = 0.041). The strongest effects were observed in the right superior longitudinal fasciculus (ρ = +0.915) and forceps minor (ρ= +0.906), inferior fronto-occipital fasciculus (ρ= +0.909). Three tracts showed significant negative correlations, and three other tracts showed no significant trends. Examples of tract orientation and corresponding MFC intensity can be seen in Figure 3. A significant positive relationship across all WM was also observed (Figure 4).

A positive relationship between mean MFC and mean sin⁴(θ) was found for the majority of WM tracts, and for whole-brain WM, consistent with the theoretical prediction that anisotropic field perturbation driven by myelinated axons is the primary source of field inhomogeneity contributing to the MFC in WM. Tracts with a broad angular range relative to B0, such as the forceps minor, SLF, and IFO, produced the strongest correlations (ρ > 0.9), reflecting the large sin⁴(θ) range (Figure 1). Negative correlations in the CST are interpretable as a geometric consequence rather than a biological contradiction; the CST runs predominantly parallel to B₀ (θ ≈ 0°), confining sin⁴(θ) to a narrow low range where MFC variance is likely dominated by orientation-independent sources such as macroscopic field inhomogeneities from air–tissue interfaces; visible in cortical and sulcal regions. The left cingulum (hippocampal) negative result warrants further investigation; its tortuous trajectory and partial overlap with CSF-adjacent regions may introduce partial-volume contamination into both the MFC and V₁ estimates. Several limitations should be acknowledged; the use of a population-averaged V₁ rather than subject-specific DTI means that individual variability in fibre orientation is not captured, potentially underpowering the observed correlations. Similarly, the group-mean MFC volume, while SNR-efficient, obscures inter-subject MFC variance that could itself be informative. Future work will examine these effects at an individual level. These results motivate correcting for fibre–B₀ orientation in future MFCi work and support exploration of MFC as a biomarker of myelin status in demyelinating pathologies, such as multiple sclerosis.

MFC in WM is significantly modulated by fibre–B₀ orientation in a pattern consistent with myelin-driven dipolar field anisotropy, implicating myelin's anisotropic diamagnetic susceptibility as the dominant source. These results suggest that MFCi at ultra-high field is a myelin-sensitive technique, supporting its potential utility in assessing demyelinating pathology.

Quantitative methods to map of the reversible, inhomogeneous field contribution T2’ to the effective transverse relaxation time T2* have gained particular interest over the recent years. Especially so-called χ-separation methods are promising to distinguish tissues with para- and diamagnetic susceptibility and may be helpful to better characterize brain tissue [1]. While χ-separation using only T2* is possible, accurate knowledge of T2’ is preferable. To this end, T2 and T2* need to be fitted first. Most often, this is done based on separately obtained and qualitatively very different MRI data, like multi-echo spin echo data (ME-SE for T2 fit) and multi-echo gradient echo data (ME-GRE for T2* fit). However, such vastly different pulse sequences inherently saturate tissue compartments in different ways due to altered RF patterns (e.g. 1 excitation vs. 1 excitation + several repeated refocusing pulses), slice-selective 2D ME-SE vs. volumetric 3D ME-GRE sequences, or even just different repetition times and flip angles. This may result in varied “visibility” of tissue compartments in a voxel’s signal and can ultimately introduce potential sequence/tissue-bias in the resulting T2, T2* and T2' map. In this work, we propose to fit T2 and T2* from a single ME-GRE scan including several volume measurements rapidly acquired using a segmented 3D-EPI approach with identical RF pulses (flip angle) but varied RF spoil increments Δφ to make the signal amplitude and phase sensitive to T2. This allows to fit T2 to the complex GRE data. Previously, such T2-mapping methods [2,3] have – to the best of our knowledge – only been applied to single-echo data. We demonstrate that multi-dimensional data helps to fit several parameters of interest simultaneously. Furthermore, T2* bias is avoided in the subsequent T2 estimation, if the mono-exponentially fitted TE=0 magnitudes are used. Because of the prolonged TR, this work also proposes adapted sequence parameters. The approach, which has reduced SAR and B1-inhomogenity constraints compared to ME-SE, is combined here with pTX excitation at 7 Tesla.

The ME-GRE signal with RF spoil increment Δφ depends on various MR tissue and sequence parameters (see signal equation of Fig. 1). Since both the T2-specific phase, Θ(|Δφ|, T1, T2) and the magnitude depends on T2 (and T1), we follow a previous magnitude- and phase-based T2 fitting approach [3] using 3 RF spoil increments with both signs. In total we acquire 24 = 6 (RF increments) x 4 (TE) magnitude and phase images. Acquisition: All scans were performed on a Siemens 7T Plus system (Siemens, Erlangen, Germany) using universal non-selective water-selective pTX excitation pulses [4] with a 32Rx/8Tx head coil (Nova Medical, Wilmington, MA). A B1 map for this pulse was acquired using the actual flip angle imaging method [5]. A multi-echo skipped-CAIPI sequence [6] was adapted to acquire an image series with varied RF spoil increments, incl. 5 seconds of initial dummy TRs to converge to each new steady-state (0.7mm iso, whole-head, CAIPI 2x2z1, EPI factor 4, TR=26ms, TA=7 min). 4 TEs = [3.4, 9.9, 14.6, 20.2]ms were chosen for a reasonable TE range within shortest possible TR. Compared to previous works [2,3], the increased TR required parameter adjustments to maintain reasonable T2-sensitization of the phase: specifically, an increased flip angle (40deg nominal) to keep the dynamic signal range high, and a 3-times wider Δφ range (-15°…15°) to account for stretched magnitude and phase curves as a function of RF spoil increment (see Fig. 2). Analysis: (1). As outlined in Fig. 1, from the 6 ME-GRE signal magnitudes per voxel, T2* and S0(|Δφ|=4.5°), S0(|Δφ|=9°), S0(|Δφ|=15°) were fit simultaneously (TE as independent variable). The extrapolated amplitudes S0 at TE=0 are free of T2* bias. (2). The background phase was eliminated pair-wise by computing the Hermitian inner product of the complex TE1=3.4ms signals at +|Δφ| and -|Δφ| RF spoil increments [3]. Note that the remaining T2-specific phase Θ(|Δφ|, T1, T2) is not dependent on T2*. (3). T2 was fit (together with amplitude and T1 estimates) based on the 3 |Δφ| using S0 (of step 1) and the background-corrected phase Θ (of step 2) using an extended-phase-graph (EPG) fitting routine (see [3]). (4). T2’ was finally obtained from 1/T2’ = 1/T2* -1/T2 using the relaxation times of step 1 and 3.

Fig. 3 shows an example slice of T2 (of step 3), of T2* (of step 1) and of T2’ (of step 4).

Obtained relaxation time maps show plausible and homogeneous values in CSF, WM, GM and deep GM across the brain. As indicated by Fig. 1, the potential of the data is not yet fully exploited: the multi-echo data also contain frequency information (ω) that is going to be used in the future for QSM and/or χ-separation.

As a proof-of-concept, this work has shown that high-resolution whole-brain T2’ (T2 and T2*) maps can be obtained from a single ME-GRE 3D-EPI image series with varied RF spoil increments at 7 Tesla.

Quantitative Susceptibility Mapping (QSM) at high and ultra-high field presents strong benefits due to an enhancement of the susceptibility effect[1]. While R2* is expected to increase linearly with field strength[2], susceptibility values are theoretically field-independent. However, slightly lower susceptibility values at 7T compared to 3T have been reported, potentially reflecting acquisition-related biases[3]. Furthermore, QSM strongly depends on echo time, limiting inter-field reproducibility when identical parameters are used[4]. Stronger B0 inhomogeneities at 7 and 11.7T further complicate cross-field comparison. Here, we independently optimized acquisition protocols at each field strength and compared R2* and QSM at 3T and 7T in 10 healthy subjects using multiple sequences and resolutions with identical reconstruction pipelines. In addition, preliminary data from one subject acquired at 11.7T on the Magnetom Iseult are also presented. Normative values were derived in the basal ganglia using a combined deep learning segmentation approach.

Ten healthy volunteers (mean age 30±5 years, 5 women/5 men) were scanned on Siemens Healthineers 3T Magnetom Cima.X, 7T Magnetom Terra.X and 11.7T Magnetom Iseult[5] systems (software: XA61, XA60, VE12U respectively). At 7 and 11.7T, pTx mode with non-selective Universal Pulses (UP)[6] was used to ensure homogeneous flip angle distribution, maximizing SNR uniformity across the brain. A 64-channel receive-only head coil (Siemens Healthineers) was used at 3T, an 8/32-channel pTx coil (NovaMedical) at 7T and a custom 8/32-channel pTx head coil at 11.7T[7]. Whole-brain 3D Multi-Echo GRE, VIBE and segmented EPI[8] sequences were acquired at 1.0 and 0.6 mm isotropic resolutions (Table 1). All subjects underwent two MRI sessions (V1 and V2, 7-day interval) for scan-rescan reproducibility assessment at 3 and 7T but only one session at 11.7T. Echo images were combined using root mean square (RMS). Brain masks from BET (FSL) on 3D GRE V1 were registered to all sequences and sessions using ANTs[9]. R2* was fitted using nonlinear regression. For QSM, phase evolution across successive TEs was first fitted to estimate the local frequency shift map, providing TE-robust field estimation, followed by background field removal and dipole inversion using MEDI-Toolbox [10-11]. ROIs were defined using a deep learning model[12] and back-projected to subject space.

Representative maps and normative values of both R2* and QSM are shown in Figure 1. Scan-rescan reproducibility of QSM and R2* is summarized in Table 2. QSM reproducibility was excellent across all sequences and ROIs (globus pallidus (GP), putamen (Put) and caudate (Cau) nucleus) at 1mm isotropic resolution (ICC: 0.935-0.997, CoV<10%), with no significant systematic bias between sessions (Bland-Altman, all p>0.05). At 0.6mm, QSM reproducibility was reduced for the GRE sequence at 3T, consistent with insufficient SNR at lower field strength and higher spatial resolution. R2* reproducibility was excellent across sequences and ROIs (ICC: 0.785-0.976 at 3T, 0.660-0.974 at 7T), apart from a significant systematic bias detected for the Cau nucleus using the GRE 0.6mm sequence at 3T (bias: -0.755 Hz, p=0.010). R2* values increased progressively with field strength (GRE, 0.6mm iso). The 7T/3T ratios (2.28, 2.04 and 1.89 for GP, Put and Cau) were below the theoretical value of 7/3≈2.33, as were 11.7T/3T ratios (3.70, 3.42 and 2.97) vs 11.7/3≈3.90. At 1mm, QSM values were significantly lower at 7T for GRE (except Put), VIBE and EPI#1 (all p<0.05), while EPI#2 showed no significant difference in GP and Put. At 0.6mm, significant reductions were observed for GRE across all ROIs but not for EPI. Preliminary QSM values at 11.7T (GRE, 0.6mm iso) were 0.135, 0.055 and 0.056 ppm for the GP, Put and Cau, comparable to or slightly above 3T values.

Acquisition protocols were pushed just below PNS and SAR limits at each field strength to maximize data quality, with all sequences kept under 10 minutes for clinical translatability. R2* values showed linear evolution with field strength, confirming the expected linear relationship[2]. QSM values were slightly but significantly lower at 7T compared to 3T, suggesting either residual acquisition-related biases or a genuine field-dependence of susceptibility measurements. The 11.7T dataset will provide further insight into this question. The TE-dependency of QSM[4] was addressed by fitting phase evolution across successive TEs, reducing sensitivity to TE-dependent nonlinearities and providing more robust cross-field comparison.

This study provides normative R2* and QSM values in deep brain structures at 3T and 7T using field-optimized protocols and identical reconstruction pipelines, ensuring unbiased comparison across field strengths, sequences and resolutions. Acquisitions at 11.7T will be completed with the same cohort, paving the way for future patient studies.

Quantitative Susceptibility Imaging (QSM) is a valuable method to assess iron accumulation and ongoing inflammation. These susceptibility-based features are increasingly recognized for their diagnostic and prognostic value for multiple sclerosis (MS)[1]. To allow its application in clinical practice, it is crucial to obtain reliable maps[2], and to fully understand how the imaging parameters can affect them[3]. Previous studies have reported an echo-time dependency of QSM measurement[4–6]. However, the potential impact of the echo sampling strategy remain unclear. The goal of this study is to explore the impact of multiple strategies of echo down-sampling on QSM measurement.

This ancillary study combined datasets from two prospective studies (ClinicalTrial.gov: NCT04906941 and NCT04907487). All participants underwent scanning sessions on a 3T MRI scanner (Elition, Philips Healthcare, Netherlands), with a 32-channel head coil. We selected participants (controls and MS patients) who had an MRI with the following sequence: (i) 3D T1-weigthed magnetization-prepared rapid gradient-echo), (ii) 3D Fluid attenuated inversion recovery (FLAIR), and (iii) 3D gradient-echo (GRE) susceptibility-weighted images (SWI) using 12 monopolar echoes. Details of imaging parameters are presented Fig. 1. Three strategies of echo downsampling were evaluated by keeping specific set of 6 echoes: (i) firsts, removing last 6 echoes; (ii) odd, keeping an echo out of two starting at the first one; and (iii) even, keeping an echo out of two starting at the second one. SWI sequence acquired using 12 echoes was considered as the reference. QSM maps were reconstructed using the SEPIA software[7]. Brain were extracted using FSL-Bet[8]. Phase unwrapping was done with Laplacian-based method[9], and echo combination with optimum weights[10]. The background field removal was performed using a PDF method[11]. The dipole inversion was solved by the MEDI method[12][13] with the regularisation parameter lambda set to 1000. Image analysis was performed in the T1-weighted native space by registering QSM maps with ANTS. Region-of-interest were extracted using FIRST (FMRIB Software Library), Freesurfer, and DeepLesionBrain (VolBrain). For each ROI, the mean and the standard deviation (SD) of QSM signal were extracted. We compared the mean value of each down-sampling strategies with the reference map using paired t-tests with Bonferroni correction. Error map, bias, normalised absolute error (NMAE), structural similarity index optimized for QSM (XSIM)[14] and intra-class correlation coefficient (ICC) were also computed.

89 participants were included (Female=60) with a mean age of 41±13 years old, including 81 patients with MS. Fig. 2 shows an example of QSM maps generated with the three echo down-sampling strategies, and the corresponding error and XSIM maps. Fig. 3 presents the mean and standard deviation of regional QSM values. Down-sampled reconstructions have significantly different mean value over most of the ROI in comparison with the reference map. Fig. 4 shows four comparison metrics (Bias, NMAE, XSIM and ICC) computed for each ROI and downsampled QSM maps. Caudate nuclei exhibit the highest bias (Firsts: -4.2±4.1ppb, Odd: -4.4±2.9ppb, Even: 5.8±3.5ppb). Thalami present the highest NMAE (Firsts: 0.72±0.17, Odd: 0.49±0.11, Even: 0.67±0.13) and increase of intra-ROI variability, associated with the lowest XSIM (First: 0.54±0.07, Odd: 0.76±0.05, Even: 0.57±0.06). The reproducibility between the reference and the down-sampled maps is good to excellent except for normal-appearing white-matter (ICC NAWM Even=0.68). Lesion and putamen are regions the most robust in term of reproducibility among subjects and structural similarity (ICC>0.94, XSIM>0.68).

By down-sampling multi-echo SWI data, we evaluated how the number and distribution of echoes influence QSM signal reliability, reproducibility, and structural fidelity across different brain regions. Reduced or non-optimally sampled echo sets introduce measurable bias and structural deviations. Using only six echoes with an interval of one echo resulted in greater errors than using an interval of two, suggesting that the first six echoes do not provide sufficient temporal coverage. The results also highlight the importance of the first echo on the QSM measurements. Despite this, overall reproducibility remained good to excellent (ICC>0.75) except in NAWM. These results confirm the echo-time dependency of QSM reported in previous studies and emphasize that balanced inclusion of both early and late echoes is crucial to minimize bias and preserve image fidelity.

These findings demonstrate that the echo sampling strategy chosen impacts brain regions differently. Our results, obtained on a large cohort of patients, highlight that limited echo coverage leads to regional bias and reduced accuracy. Optimizing echo sampling is essential for reliable QSM measurements in research and clinical applications.

Quantitative Susceptibility Mapping (QSM) is increasingly used in multiple sclerosis (MS) to assess iron-related pathology and susceptibility-based biomarkers such as paramagnetic rim lesions [1]. However, the influence of gadolinium-based contrast agents (GBCAs) on QSM measurements remains insufficiently characterized, despite frequent post-contrast acquisition of susceptibility sequences in clinical workflows. This study investigates how gadolinium affects QSM measurements in MS patients undergoing contrast-enhanced MRIs, analyzing QSM values before and after GBCAs injection while considering timing and previous injections. As susceptibility-based imaging gains diagnostic and prognostic importance in MS, understanding potential GBCA-related bias is essential for reliable interpretation.

In this prospective single-center study (ClinicalTrial.gov: NCT04906941), patients with MS underwent brain magnetic resonance imaging including identical QSM acquisitions before and after GBCA administration (Figure 1). QSM raw data were acquired using a 3D-FFE sequence with the following parameters: FOV: 240*240*176 mm3, Voxel size = 1.0*1.2*1.0 mm3 reconstructed to 1.0*1.0*1.0 mm3. Elliptical scan, TR = 41 ms, TEs = 3.1 ms to 36.1 ms with a ΔTE of 3 ms, Bandwidth per pixel = 913.7 Hz. For susceptibility maps reconstruction, the local field map was calculated by non-linear fitting the complex gradient echo signal over echo times, then background field removal was performed using the pre-conditioned Laplacian boundary value method (LBV) and the inversion was solved using the L1-MEDI method [2]. Automated anatomical segmentation was performed using FreeSurfer and FSL-FIRST to extract regional QSM values in white matter, cortex, deep gray matter, and fluid-attenuated inversion recovery lesions. Pre- and post-contrast QSM values were compared using paired statistical testing with Bonferroni correction, intra-class correlation coefficient (ICC), and Bland-Altman plots. Linear mixed-effects modeling assessed the influence of GBCA administration while adjusting for age and sex. Associations with GBCA dose, injection timing, and prior GBCA exposure were also explored.

Fifty-five participants (38 women, mean age ± standard deviation = 39.8±12.8 years old) were included after excluding scans with incomplete data or significant motion artifacts. GBCA administration resulted in significant reductions in QSM values across all examined brain regions (all p<0.01, Figure 2, Table 1). In global white matter, mean QSM values decreased from −4.87±0.35 ppb pre-contrast to −7.92±0.36 ppb post-contrast (p<0.001). Similar reductions were observed in cortical and deep gray matter regions, with the largest susceptibility decrease detected in deep gray matter structures. White matter lesions showed smaller average changes (5.63±0.91 vs 4.36±1.01 ppb, p=0.003) but greater inter-subject variability. ICC indicated poor-to-moderate agreement in white matter (ICC=0.41; 95% confidence interval [CI]: 0.17, 0.61) and cortex (ICC=0.71; 95% CI: 0.55, 0.82), while excellent agreement was observed in lesions (ICC=0.92; 95% CI: 0.87, 0.96). Bland–Altman analysis further demonstrated that the observed post-contrast differences exceeded expected scan–rescan variability in several key regions, particularly white matter and deep gray matter. The magnitude of QSM reduction correlated with administered GBCA dose (Figure 3), indicating a dose-dependent effect. In contrast, no significant association was found between QSM change and body weight or prior GBCA exposure. In the linear mixed-effects modeling, pre-contrast scans showed significantly higher QSM values than post-contrast scans (β=3.71±0.26, p<0.0001), independent of age and sex.

These findings demonstrate that GBCAs induce a systematic susceptibility bias affecting QSM measurements in MS. The regional heterogeneity observed may reflect differences in vascularization, baseline susceptibility, and lesion composition. Because QSM is increasingly used to characterize iron accumulation and chronic active lesions, post-contrast acquisitions may confound interpretation in both clinical and research settings. Although susceptibility-weighted imaging was not directly assessed, the shared dependence on phase information suggests that similar effects may occur in susceptibility-weighted imaging–based biomarkers.

GBCA administration significantly alters QSM measurements in patients with MS, introducing a systematic susceptibility bias that may confound susceptibility-based biomarkers. These findings support standardizing magnetic resonance imaging protocols to acquire QSM before contrast administration, particularly in studies evaluating iron deposition or paramagnetic rim lesions.

The 2024 revision of the McDonald criteria highlights the central vein sign (CVS) and paramagnetic rim lesion (PRL) as key imaging features for multiple sclerosis (MS) diagnosis [1]. Reliable detection of these imaging features depends on susceptibility-weighted imaging (SWI) enhancing intraparenchymal veins visibility on magnitude images and iron deposition on phase images [2]. However, the optimal susceptibility-weighted imaging (SWI) acquisition (parameters and sequences) for reliable detection remains uncertain. In this study, we compared four SWI sequences to identify the most sensitive and clinically feasible approach for reliable CVS and PRL detection, supporting their integration into routine MS diagnostic workflows.

In this prospective cross-sectional study (ClinicalTrials.gov: NCT04705870), participants with MS diagnosed according to the 2017 McDonald criteria underwent a standardized protocol on a 3 Tesla MRI scanner (Elition, Philips Healthcare, Netherlands). Four SWI sequences were acquired: optimized EPI-based SWI (SWI-EPIisovoxel), gradient-echo SWI (SWI-GRE), resolution-matched EPI-SWI (SWI-EPIanisovoxe), and repetition-time–matched EPI-SWI (SWI-EPIshortTR). Phase images were reconstructed from the complex signal and processed using vendor-provided high-pass filtering to generate susceptibility-weighted phase images. For each participant, FLAIR images were reviewed to identify up to five lesions per anatomical region meeting predefined criteria (Figure 1)[3] , with an overall cap of 20 lesions per subject. A standardized report form recorded the number of selected lesions, CVS-positive lesions, PRL, and lesions with multiple veins. Subjective ratings included image quality, artefact severity, and diagnostic confidence for CVS and PRL (Table 1). Paired comparisons and linear mixed-effects models were performed. Subjective ratings were analysed using the Friedman test, followed by pairwise Wilcoxon signed-rank tests with Bonferroni correction for multiple comparisons.

Fifty-six participants (mean age 38.2±11.3 years; 29 men) with 453 lesions were analyzed (180 periventricular, 50 juxta-cortical, 185 deep white matter and 38 posterior fossa). SWI-EPIisovoxel detected a higher proportion of CVS-positive lesions compared with SWI-GRE (68.6% vs 41.3%, p<0.001, Figure 2). SWI-EPIanisovoxel and SWI-EPIshortTR also outperformed SWI-GRE (67.4%, p<0.001; 65.5%, p<0.001; respectively). No differences were observed between EPI-based sequences (p=1.0). PRL detection was comparable across sequences (12.9%–17.3%, Figure 3). Subjective evaluation showed significantly higher diagnostic confidence for CVS with EPI-based sequences compared with SWI-GRE (χ² = 64.28, p<0.001). Pairwise analyses confirmed improvements for SWI-EPIisovoxel (p = 1.2 × 10⁻⁷), SWI-EPIanisovoxel (p = 4.1 × 10⁻⁶), and SWI-EPIshortTR (p = 1.2 × 10⁻⁴). Diagnostic confidence for PRL also differed across sequences (χ² = 21.34, p < 0.0001), with higher confidence for SWI-EPIanisovoxel and SWI-EPIshortTR compared with SWI-GRE (p = 0.001 and p = 0.003, respectively). No differences were observed between EPI-based sequencesfor CVS and PRL detection. Overall image quality was rated high across all sequences, with median scores of 5 for each. Pairwise comparisons revealed that only SWI-EPIanisovoxel and SWI-EPIshortTR achieved better scores than SWI-GRE (p=0.007 and p=0.003, respectively). Artifact ratings were low and comparable across all acquisitions (median = 4, range 2–5, p=0.25).

Our study demonstrated an improved visualization of the CVS using SWI-EPI sequences compared to a gradient-echo SWI images in participants with MS. Furthermore, sequences specifically optimized for CVS detection also provided superior visualization. Following the last revision of the McDonald criteria, reliable CVS assessment is particularly relevant, as it may refine risk stratification in patients with incidentally discovered white matter lesions by helping to distinguish demyelinating pathology from alternative causes. CVS assessment is also highly relevant in the differentiation of MS from vascular-related white matter hyperintensities. Regarding PRLs, our results indicate that sequence optimization had limited impact on their detection. This finding suggests that PRL visibility may be less dependent on acquisition strategy than CVS detection, and more strongly influenced by lesion biology and disease stage.

Our study demonstrates the superiority of EPI-based SWI over SWI-GRE for CVS visualization and reader confidence compared, supporting their use for MRI-based MS diagnosis. Given the increasing weight of CVS in the revised 2024 McDonald criteria, the routine implementation of SWI-EPI sequences should be strongly considered for MS diagnosis and monitoring in clinical practice.