Comparison of multi echo T2 relaxation and steady state approaches for myelin imaging in the central nervous system

The traditional approach for measuring myelin-associated water with quantitative magnetic resonance imaging (MRI) uses multi-echo T2 relaxation data to calculate the myelin water fraction (MWF). A fundamentally different approach, abbreviated “mcDESPOT”, uses a more efficient steady-state acquisition to generate an equivalent metric (fM). Although previous studies have demonstrated inherent instability and bias in the complex mcDESPOT analysis procedure, fM has often been used as a surrogate for MWF. We produced and compared multivariate atlases of MWF and fM in healthy human brain and cervical spinal cord (available online) and compared their ability to detect multiple sclerosis pathology. A significant bias was found in all regions (p < 10–5), albeit reversed for spinal cord (fM-MWF =  − 3.4%) compared to brain (+ 6.2%). MWF and fM followed an approximately linear relationship for regions with MWF <  ~ 10%. For MWF >  ~ 10%, the relationship broke down and fM no longer increased in tandem with MWF. For multiple sclerosis patients, MWF and fM Z score maps showed overlapping areas of low Z score and similar trends between patients and brain regions, although those of fM generally had greater spatial extent and magnitude of severity. These results will guide future choice of myelin-sensitive quantitative MRI and improve interpretation of studies using either myelin imaging approach.

The traditional approach for measuring myelin-associated water with quantitative magnetic resonance imaging (MRI) uses multi-echo T 2 relaxation data to calculate the myelin water fraction (MWF). A fundamentally different approach, abbreviated "mcDESPOT", uses a more efficient steadystate acquisition to generate an equivalent metric (f M ). Although previous studies have demonstrated inherent instability and bias in the complex mcDESPOT analysis procedure, f M has often been used as a surrogate for MWF. We produced and compared multivariate atlases of MWF and f M in healthy human brain and cervical spinal cord (available online) and compared their ability to detect multiple sclerosis pathology. A significant bias was found in all regions (p < 10 -5 ), albeit reversed for spinal cord (f M -MWF = − 3.4%) compared to brain (+ 6.2%). MWF and f M followed an approximately linear relationship for regions with MWF < ~ 10%. For MWF > ~ 10%, the relationship broke down and f M no longer increased in tandem with MWF. For multiple sclerosis patients, MWF and f M Z score maps showed overlapping areas of low Z score and similar trends between patients and brain regions, although those of f M generally had greater spatial extent and magnitude of severity. These results will guide future choice of myelin-sensitive quantitative MRI and improve interpretation of studies using either myelin imaging approach.
Quantitative magnetic resonance imaging (MRI) techniques are developed with the aim of being specific to certain features of tissue microstructure. In the case of myelin water imaging, the water trapped between myelin bilayers can be used as a biomarker for myelin content.
The original method for myelin water imaging was a multi-echo spin-echo T 2 relaxation acquisition, which can quantify signal contributions from different characteristic T 2 relaxation times, such as those associated with water trapped between myelin lipid bilayers (T 2 < 40 ms at 3 T) or water within intra-and extra-cellular spaces (40 ms < T 2 < 200 ms) 1 . The myelin water fraction (MWF) metric can then be calculated as the signal contribution from T 2 relaxation times associated with myelin water divided by the total signal. Although this model neglects the presence of non-aqueous protons and the effects of magnetization exchange between water pools 2,3 , the MWF has been shown to correlate well with quantitative histopathologic measures of myelin density [4][5][6] . A multi-echo 3D GRAdient-echo and Spin-Echo (GRASE) sequence can be used to reduce acquisition time by a factor of 3 with a negligible reduction in the resulting MWF map quality compared to a multi-echo spin-echo sequence 7,8 .
The value of an in-vivo myelin biomarker has motivated the development of alternative approaches to myelin imaging. One commonly used alternative is the multi-component Driven-Equilibrium Single-Pulse Observation of T 1 and T 2 (mcDESPOT) technique, which is based on a series of steady-state acquisitions 9 . Spoiled gradient recalled echo (SPGR), inversion-recovery-prepared SPGR (IRSPGR) and balanced steady-state free precession (bSSFP) images are acquired with a range of flip angles (FA), from which T 1 and T 2 can be estimated. The mcDESPOT analysis framework requires fitting the data to a large number of parameters simultaneously, • create atlases of normative reference MWF and f M values for both brain and cervical spinal cord; • use atlas-based methods to compare myelin imaging approaches with minimal influence from biological variation and measurement noise; • improve characterisation of the relationship between MWF and f M , by comparing values across a diverse range of tissue, including healthy and diseased brain and spinal cord.

Results
Five of the 28 participants had to be excluded from the MWF and f M spinal cord atlases, due to imaging artefacts that rendered their GRASE or mcDESPOT data unusable.  Figure 1A shows f M plotted against MWF, for each ROI value. The Bland-Altman plot in Fig. 1B shows the difference between each MWF and f M ROI value plotted against their mean. There was an approximately linear relationship between MWF and f M for brain regions (circle data points in Fig. 1A) with MWF < ~ 10%. The linear MWF-f M relationship did not hold in the spinal cord, or in brain regions with MWF > ~ 10%. The non-linearity between f M and MWF appears qualitatively consistent across participants, with a tight spread of data points following the same distinct pattern. The intra-subject, intra-ROI variability of each metric, indicated by standard deviation uncertainty bars in Fig. 1A, is small compared to the range of f M and MWF values and could not account for the non-linear relationship.

Mean atlas ROI values.
In the spinal cord, MWF spanned a large range of values (12-30%) whereas f M had little variation (16-23%). A systematic bias exists between MWF and f M values, but with opposite sign for brain (mean WM&GM f M -MWF = 6.2%) compared to spinal cord (mean whole cord f M -MWF = − 3.4%).

Templates and atlases.
Representative slices of the anatomical templates, T 1 -weighted (T 1 w) for brain and T 2 *-weighted (T 2 *w) for spinal cord, voxel-wise mean MWF and f M atlases, and voxel-wise SD MWF and f M atlases are displayed in Fig. 2A Fig. 3B, show widespread regions of low Z score for both MWF and f M in both brain and spinal cord. In Fig. 3, histograms for brain and cord of each patient show similar distributions of Z score values for MWF and f M , with the largest differences visible in the extremes of the Z score range for PPMS. For RRMS, a slight shift towards lower values is apparent for the MWF Z score histogram, compared to f M . For PPMS, the MWF Z score maps are generally less extreme in terms of both spatial extent and statistical magnitude compared to those of f M . While PPMS Z score maps suggest that f M has greater sensitivity to the presence of pathological tissue changes, RRMS Z score histograms suggest that MWF may better identify subtle changes outside regions of more-extreme pathology. Regions of deficient MWF Z score are generally highlighted on the corresponding f M Z score map, although this congruity appears to be more consistent in brain.

Discussion
The atlas-based methods employed here provided an effective means of reducing biological variations and measurement noise present within and between individual participant quantitative metric maps. In addition to facilitating the creation of Z score maps, atlas-based methods provide increased statistical power for characterizing discrepancies between the quantitative myelin fraction metrics obtained with GRASE and mcDESPOT. The smooth, "high SNR" appearance of the atlases in Fig. 2 provide clear depictions of representative values expected Table 1. Mean and standard deviation (SD) of voxels within each region of interest (ROI) for the mean myelin water fraction atlases from multi-echo T 2 relaxation (MWF) and steady-state (f M ) approaches. ROI volumes are also provided. A paired, two-tailed t test found significantly different MWF and f M ROI values (p < 10 -5 ) across healthy control participants, for all ROIs. Brain ROIs include all white matter (WM), all gray matter (GM) and combined WM&GM from T 1 w image segmentations, along with 9 additional WM ROIs (all JHU white matter labels combined, genu of corpus callosum, splenium of corpus callosum, whole corpus callosum (CC), posterior internal capsule, and frontal, occipital, parietal, and temporal lobes masked to WM), and 4 additional GM ROIs (cortical GM, caudate, thalamus, and putamen). Spinal cord ROIs include the whole cord, WM, GM, dorsal column, and lateral funiculi. www.nature.com/scientificreports/ from each imaging modality. They can also serve as a reference for how MWF and f M values are expected to differ throughout commonly investigated regions in the central nervous system. The MWF and f M atlases are openly available and provide normative reference data to increase statistical power for future studies with limited ability to recruit and collect data from healthy controls. Although previous studies have compared MWF and f M in the brain, this is the first comparison of MWF and f M values in both the brain and spinal cord. Our combined analysis yielded some consequential results, especially when interpreting spinal cord results in concert with those from brain. First, the absolute difference in magnitude between myelin-related metric values produced by multi-echo T 2 relaxation and steady-state approaches is much smaller in spinal cord than in brain. Despite having more-similar magnitudes than in brain, MWF and f M in the spinal cord bear almost no correspondence. Secondly, the average bias between GRASE and mcDESPOT is reversed in spinal cord (mean whole cord f M -MWF = − 3.4%) compared to brain (mean WM&GM f M -MWF = 6.2%). The inversion of the MWF versus f M bias between brain and spinal cord is clearly depicted by the individual participant mean ROI values plotted in Fig. 1A, where most brain and spinal cord data points lie on opposite sides of the line of unity.
Although comparisons in human brain have shown significant, moderately strong correlations between MWF and f M 26 , the same linear relationship does not hold in the spinal cord. Figure 1A suggests that this relationship  www.nature.com/scientificreports/ actually begins to break down in brain regions as well. Specifically, MWF and f M do not appear to be correlated in regions with MWF > ~ 10%, where f M no longer increases in tandem with MWF. In hindsight, this is visible in some brain regions examined by previous studies 26,34 , but the magnitude and extent of the incongruity between the two methods is made clear by our unified, direct comparison with both brain and spinal cord data. Our atlas ROI results (Table 1) show that MWF is lower than f M throughout all investigated brain ROIs, in agreement with previous studies that compared values in the brain [24][25][26] . Lack of contrast in f M between WM ROIs, as demonstrated by Table 1 values and qualitatively by mean f M atlases in Fig. 2, also matches expectations based on previous brain studies with mcDESPOT [24][25][26] .
Previous work has provided evidence that myelin content within human brain WM is heterogeneous. Aboitiz et al. studied the corpus callosum in post-mortem human brain using the Loyez histological stain for myelin and electron microscopy 35 . They found that the genu had a smaller median axon diameter (0.6 μm) compared to the splenium (1.0 μm), which is expected to correspond with lower oligodendrocyte proliferation and myelin content 35 . They also found a high percentage of unmyelinated fibers in the genu (16%) compared to the rest of the corpus callosum (< 5%) 35 . Both of these findings support the expectation of lower myelin content in the genu, compared to the splenium, which we found reflected in our results for MWF but not for f M . In more recent studies using histological staining markers for myelin, the optical density of Luxol fast blue showed regional variation by factors > 2 within non-lesional WM, albeit in MS brain tissue where pathological changes could increase stain contrast 5,6 . Bürgel et al. demonstrated strong regional contrast in WM myelination when they used a modified Heidenhain-Woelcke stain 36 to distinguish major tracts (including the corticospinal tract, optic and acoustic radiations, fornix, cingulum, corpus callosum, superior longitudinal, superior and inferior occipito-frontal and uncinate fascicles) in 10 post-mortem brains by their relative degrees of myelination 37 . Therefore, the low contrast of f M within WM is likely not representative of myelin content but rather could relate to a confounding factor for   28 . In our study, WM accounted for 78% of the spinal cord volume, compared to 43% of brain tissue volume. As a result, higher mean MWF and f M can generally be expected across the whole spinal cord. Although this does not explain higher MWF in both WM and GM individually, previous studies have reported approximately 2.5 times higher myelin content in the cord, compared to the cerebrum, of rat nervous system 39 , providing a precedent for a similar pattern to exist in human tissue.
There are many possible explanations for the low f M in the spinal cord, compared to expectations from the brain. Like all quantitative MRI techniques, GRASE and mcDESPOT each have their own set of confounds, which could affect the results more or less in the spinal cord, causing the relationship between MWF and f M to change.
GRASE is expected to have some magnetization transfer effects due to repeated refocusing pulses throughout the echo train 40 . For mcDESPOT, the short, high-amplitude excitation pulses used to acquire SPGR and bSSFP www.nature.com/scientificreports/ data minimize undesirable finite radio-frequency pulse effects but increase the effects of magnetization transfer. By sensitizing the signal to macromolecular protons present in biological tissue, magnetization transfer can have a profound effect on estimates of T 1 and T 2 relaxation values [41][42][43] . Modified versions of mcDESPOT have been introduced with additional model parameters in an attempt to quantify and correct for magnetization transfer 44 . Upon investigation, Zhang et al. confirmed that removal of magnetization transfer has a strong effect on bSSFP signal curves. However, values for f M remained much higher than MWF, suggesting that magnetization transfer does not drive the MWF-f M difference in brain. The fact that signal changes did not propagate into the expected parameter changes helped motivate subsequent investigation into the mcDESPOT analysis procedure. The GRASE MWF is generally underestimated in regions with very low signal contribution from myelin water, such as cortical GM brain regions, where there is insufficient signal-to-noise for accurate myelin water quantification 45 . The multi-component T 2 analysis used with GRASE ignores the effects of T 1 relaxation. Using a short repetition time (TR) with GRASE could therefore artificially increase MWF values, due to additional T 1 weighting. However, our TR of 1073 ms is not expected to have a large influence on MWF and does not explain the discrepancy with f M cord values 46 . Recent studies have shown that the MWF will decrease if the inter-compartmental exchange rate between myelin water and intra/extra-cellular water increases 40,47 . Therefore, GRASE MWF values may be underestimated, especially in regions with small axons and thin myelin sheaths, where an increase in inter-compartmental exchange can be expected 48 . In comparison, mcDESPOT should be immune to this effect, assuming the analysis includes and accurately estimates the myelin water residence time parameter that is intended to account for exchange. However, omitting exchange from the mcDESPOT model so it is not accounted for, as with GRASE, does not remove the discrepancy between MWF and f M values 25 .
The caveat that the residence time must be accurately estimated becomes problematic in the context of recent studies evaluating the mcDESPOT analysis. Degeneracy of parameter solutions in mcDESPOT, a fundamental problem afflicting many MRI microstructure mapping techniques 49 , was first shown by Lankford and Does studying mcDESPOT precision with Cramér-Rao lower bounds 12 . Imprecision of mcDESPOT solutions has been further demonstrated by Zhang et al. and Bouhrara et al., who both showed flat residual plots of mcDESPOT parameter fits 14,25 . This can result in substantial instability in metric values, especially for moderate or low signalto-noise data 14 . They also showed that the instability could be reduced by removing fit parameters, specifically by ignoring the effects of exchange 25 .
A recent investigation by West et al. further showed that mcDESPOT solutions are inaccurate and imprecise when exchange is included in the model 15 . As in previous work, they found that ignoring exchange improved the stability of the analysis, but at the cost of significantly biased results 15 . West et al. also found that changes in mcDESPOT analysis results were non-linear with respect to changes in their true values, and that changing a single parameter could induce marked changes in other, static parameters 15 . For example, a change in the T 1 or T 2 value of intra-/extra-cellular water can result in a large change in f M results, even when the true value of f M has not changed 15 . The notable sensitivity of f M to pathological tissue differences could be in part explained by certain pathology-related changes to T 1 or T 2 inducing a large, non-linear response in the f M results 15 .
West et al. suggested that unpredictable bias in mcDESPOT results can be at least partially mitigated by the use of consistent acquisition and analysis parameters 15 . However, our results suggest that even with consistent acquisition and analysis conditions, changes in signal characteristics due to the presence of pathology (multiple sclerosis) or unique tissue structures (spinal cord) propagate into f M values in a way that does not consistently reflect MWF values.
The anatomical structure of the spinal cord likely plays an important role in our observed breakdown of the MWF-f M relationship as well as the lack of contrast observed in our f M spinal cord results. The non-linear behaviour of the mcDESPOT analysis procedure could lead to biased results that do not necessarily reflect the true underlying values. Spinal cord WM is similar to that of corpus callosum in the brain, and cord GM is similar to putamen 50 . Therefore, based on contrast between brain values, we can expect that MWF will have more contrast between these regions than f M . Another consideration is the presence of a low f M boundary around the cord, larger than that visible for MWF, which is likely related to partial volume effects with cerebrospinal fluid. Partial volume effects could indicate imperfections in the motion correction alignment of each image volume in mcDESPOT data sets. Misalignment from motion relates to the fact that mcDESPOT acquires central k-space data separately for each image volume, spread across the entire acquisition time, whereas GRASE acquires central k-space data only once. Imperfect alignment of mcDESPOT image volumes would lead to blurred metric maps and reduced contrast between ROIs inside the cord.
MWF and f M can be used to identify pathology on an individual basis in the brain or spinal cord [51][52][53] . Here, we demonstrated this using Z score maps, shown in Fig. 3, which showed greater spatial extent and magnitude of severity for mcDESPOT f M Z scores, compared to those of GRASE MWF.
The high sensitivity of mcDESPOT to pathological differences has been previously demonstrated in the brain by O'Muircheartaigh et al. 26 . Although there is no definitive explanation, the high sensitivity of mcDESPOT is likely due to multiple compounding factors. Relatively low inter-participant f M variability, as evidenced by the low values of f M standard deviation between participants (Fig. 2), will tend to amplify Z score values. Another influence may be reduced specificity to myelin, which would grant f M sensitivity to a variety of pathological changes compared to MWF. Strong correlation (median ρ = 0.8) between f M and quantitative T 1 results from mcDESPOT 26 provides evidence suggesting that f M is closely related to, if not driven by, relaxation values. Considering this, f M may not provide enough additional independent, myelin-specific information, compared to the quantitative T 1 metric from DESPOT1-HIFI 20 , to justify the increased acquisition time to acquire a full mcDESPOT data set.
Despite their different acquisitions, analyses, and confounding factors, MWF and f M Z score histogram plots in Fig. 3 show similarity between participants and anatomical structures. Similarity between Z scores and the higher sensitivity of f M to pathological tissue changes support the interpretation that MWF and f M are both sensitive to myelin content but that f M has reduced specificity due the unpredictable, non-linear analysis response to signal www.nature.com/scientificreports/ changes. For specificity to myelin, recent technical developments allow acquisition of multi-echo T 2 relaxation data with compressed sensing acceleration, which can provide improved spatial and parametric resolution in less acquisition time, compared to the GRASE sequence presented here 54 . When possible within time constraints, studies should ideally include multiple myelin-sensitive metrics. Additional metrics incorporated in a multimodal imaging approach can provide supplemental information as well as increase confidence in the findings.
In conclusion, we have produced and provided normative reference MWF and f M atlases for both brain and cervical spinal cord. In comparing these atlases, we found that MWF and f M follow an approximately linear relationship for regions with MWF < ~ 10%, above which the relationship breaks down and f M no longer increases in tandem with MWF. In the context of demyelinating disease pathology, Z score maps from f M showed greater spatial extent and magnitude of severity. The high sensitivity of f M is a benefit for detecting pathological changes but is likely related to the mcDESPOT analysis procedure propagating MR signal changes into the f M results non-linearly and greatly limits the ability to attribute those differences solely to changes in myelin. Despite the fundamental differences in these two techniques, MWF and f M Z score maps showed overlapping areas of low Z score and similar trends between patients and brain and spinal cord regions. Ultimately, these results will guide future choice of myelin-sensitive quantitative MRI and help improve retrospective interpretation of clinical studies using multi-echo T 2 relaxation or steady-state myelin imaging measures in the context of healthy or diseased central nervous system.

Ethics. The study operated under ethics approval from the University of British Columbia Clinical Research
Ethics Board. All participants provided written informed consent. All methods were performed in accordance with the relevant guidelines and regulations.
Participants. Data were collected from 28 healthy participants without a history of brain or spinal cord disease or injury (10 male and 18 female, mean age 46 years, range 22-65 years). To facilitate comparison of MWF and f M in the context of pathology, data were also collected from two participants with clinically definite multiple sclerosis fulfilling the 2017 revised MacDonald criteria for diagnosis 55 : one participant with RRMS (female, age 26 years, expanded disability status scale (EDSS 56 ) 1.5, disease duration 3 years) and one with PPMS (female, age 64 years, EDSS 8.5, disease duration 29 years).

MRI. Imaging was performed at the UBC MRI Research Centre on a 3.0T MRI scanner (Achieva, Philips
Medical Systems, Best, The Netherlands).
Cervical spinal cord imaging used a 6-channel spine coil, with all imaging FOV centred at the C2/C3 level of the cervical spinal cord and angulation aligned with the cord.
An axial 2D multi-slice T 2 *-weighted multi-echo gradient echo image (T 2 *w) was acquired with TR = 815 ms, TE = 6.5 ms, ΔTE = 8. The T 1 w brain and T 2 *w cord images were acquired for their strong white/gray matter contrast, to facilitate alignment of anatomical structures between participants. All acquisitions were performed without cardiac or respiratory triggering. Brain and cord acquisitions for GRASE and mcDESPOT differed slightly due to differences in optimisation of parameters for each application 8,27 . www.nature.com/scientificreports/ Quantitative MRI analysis. GRASE data was analyzed using a temporally-regularized non-negative least squares algorithm with stimulated echo correction (https ://mrire searc h.med.ubc.ca/news-proje cts/myeli n-water -fract ion/), 40 T 2 relaxation delta functions logarithmically spaced from 15 to 2000 ms, 8 refocusing flip angles between 90 and 180°, and a χ 2 regularization factor of 1.02. The MWF was calculated as the fraction of total signal with 15 < T 2 < 40 ms. mcDESPOT data underwent motion correction to align the imaging volumes before the DESPOT1-HIFI code was used to obtain an estimated FA correction map from the SPGR and IRSPGR data 20 . Using the FA correction map, mcDESPOT analysis was performed to obtain the f M . The analysis was matched to that of a previous study on mcDESPOT stability, using a 3-pool model and the same stochastic-region-contraction parameter boundaries:

Image analysis and template creation. Advanced Normalization Tools software (ANTs), Spinal Cord
Toolbox (SCT), and some FMRIB Software Library (FSL) tools were used for image analysis [58][59][60] . Non-linear registrations used symmetric diffeomorphic normalization (SyN) transformations. To prevent introduction of circularity bias, no quantitative maps were involved in image registrations 61 . Workflow for the multivariate template creation process is outlined in Fig. 4.
All T 1 w, GRASE echo 1, and IRSPGR brain images underwent N4 bias field correction for low frequency intensity nonuniformities 62 . Brain segmentations were initialized by registration with the OASIS template and priors 63 then refined using Atropos n-tissue segmentation 64 . Each participant's GRASE echo 1 and IRSPGR images were masked to brain tissue before being rigidly aligned with their corresponding T 1 w image and masked again using the higher-resolution T 1 w brain segmentation.
For spinal cord images, cord segmentations were generated for each image modality using T 2 *w, GRASE echo 16, and IRSPGR images with an automated tool (sct_propseg 65,66 ) before quality control was performed manually. The cord segmentations were used to rigidly align GRASE echo 1 and IRSPGR with T 2 *w cord images. Finally, all images were masked using the higher-resolution T 2 *w cord segmentation.
Multivariate template creation was performed by iteratively co-registering the aligned and masked GRASE echo 1, IRSPGR and anatomical images (T 1 w for brain, T 2 *w for cord) using iterative rigid, affine, and SyN www.nature.com/scientificreports/ transformations. Registrations were guided by mutual information (for linear stages) and neighborhood crosscorrelation (for SyN) similarity metrics. The T 1 w and T 2 *w similarity metrics were weighted twice as heavily as those of GRASE and IRSPGR, to account for their better resolution and tissue contrast.

Region of interest definition.
A reference anatomical T 1 -weighted template was used to generate brain ROIs. This reference template was created in-house using data from 100 healthy volunteers, including those in the current study. After registration between our T 1 w template and the reference T 1 w template, JHU-ICBM-DTI-81 WM labels and MNI structural regions, optimized for the reference template using the probabilistic joint label fusion framework 67 , were transformed to our multivariate template space. Atropos n-tissue segmentation was performed on our T 1 w template to provide WM and GM segmentations 64 .
To generate spinal cord ROIs, registration was performed between our T 2 *w spinal cord template and the T 2 *-weighted SCT PAM50 template 68 , centred at the C2/C3 level of the cord, and refined using a manually defined cord GM segmentation to improve alignment of intra-cord structure. This provided probabilistic cord ROIs, which were thresholded and binarized to only include voxels with probability > 0.5, to reduce partial volume effects.
Multiple sclerosis image analysis. RRMS and PPMS images were processed the same as those of healthy participants, except they were excluded from template creation. Instead, their T 1 w and T 2 *w images were registered to the T 1 w and T 2 *w templates, and transformations were concatenated to align their MWF and f M maps in template space for comparison. Since the goal of the study was not to derive clinical conclusions, images from participants with multiple sclerosis were not segmented into lesion/non-lesion tissue.
Statistical analysis. From the healthy participant MWF maps, aligned in template space, atlases were created by calculating the voxel-wise mean and SD between participants, then cropped to exclude regions not covered by all individual participant MWF and f M FOV. For the mean MWF and f M atlases, the mean and SD within each ROI were calculated. Mean MWF and f M ROI values were compared across all healthy control participants using a paired, two-tailed t test. MWF and f M Z score maps were calculated by subtracting the mean atlas from the MS participant quantitative map, then dividing by the SD atlas. Z score maps were also masked to exclude highly variable voxels, where the coefficient of variation (calculated voxel-wise as the SD divided by the mean) was over 0.75.

Data availability
Ethics approval did not allow for making individual subjects' imaging data available. However, the anatomical templates, quantitative atlases, and ROIs created and analysed in this study have been made openly available 33 along with example code for the template creation process 69 .