Diffusion Basis Spectrum Imaging of Tissue Microstructure and Early Treatment in Multiple Sclerosis Authors: Apisit Kaewsanit, Aime Luna, Bradley Judge, Nicole Shelley, Karl Friedrichsen, Jayashree Rajamanickam, Zhude Tu, Yong Wang, Tammie L S Benzinger, Qing Wang, Matthew R Brier Published: 2026 ABSTRACT ABSTRACTObjectiveConventional MRI cannot resolve the coexisting pathological processes within multiple sclerosis tissue. Diffusion Basis Spectrum Imaging decomposes the diffusion signal into fiber, cell, and hindered ratios indexing axonal integrity, cellularity, and extracellular space, respectively. We asked how these ratios differ between early relapsing and established non‐relapsing disease, change with early treatment, and relate to disability, and secondarily, whether lesions form reproducible microstructural subtypes.MethodsThirty‐seven participants underwent multi‐shell diffusion MRI: 11 healthy controls, 15 treatment‐naïve relapsing, and 11 non‐relapsing patients. Thirteen were rescanned ~3 months after starting high‐efficacy disease‐modifying therapy. Metrics were extracted from normal‐appearing gray and white matter and lesions (core, rim, and perilesional white matter); Gaussian mixture model clustering of per‐lesion fiber, cell, and hindered ratios identified lesion subtypes.ResultsTreatment‐naïve relapsing patients showed reduced normal‐appearing white matter fiber (p = 0.04) and elevated hindered ratios (p = 0.02) versus controls. Lesions had lower fiber and cell ratios and higher hindered ratio than within‐subject normal‐appearing white matter (all p < 0.001). After treatment, lesional fiber ratio increased (p = 0.01) while lesional and gray matter cell ratios declined (both p = 0.02). Lesional cell ratio was higher in relapsing than non‐relapsing patients at the core (p = 0.03) and rim (p = 0.04) but not perilesionally; treatment‐related improvement was similar across compartments. Clustering yielded four reproducible subtypes (edematous‐hypocellular, moderately cellular, actively inflamed, structurally preserved); the inflamed subtype predominated in relapsing and the edematous‐hypocellular in non‐relapsing disease.InterpretationDiffusion basis spectrum imaging distinguishes relapsing from non‐relapsing disease by lesion‐body cellularity and detects treatment‐related changes in fiber and cell compartments. Reproducible, hypothesis‐generating lesion subtypes support its use for in vivo lesion phenotyping.Keywords: diffusion MRI, multiple sclerosis, white matter lesion Multiple sclerosis (MS) affects nearly 3 million people worldwide and remains a leading cause of non‐traumatic neurological disability in young adults [1, 2]. Its pathobiology combines acute focal inflammation that produces demyelinating lesions associated with clinical relapse with a parallel, smoldering neurodegenerative process that drives progressive axonal loss and brain atrophy, often independent of relapses [3, 4]. Emerging evidence that relapsing and progressive features coexist from the earliest disease stages has prompted a reconceptualization of MS as a biological continuum rather than a sequence of discrete clinical phases [4, 5, 6]. This shift underscores the need for imaging tools that can simultaneously resolve fiber (axon), cellular, and extracellular compartments within and across lesions, and that can capture spatial heterogeneity within the lesion itself. At the tissue level, MS lesions are heterogeneous. Histopathological studies have suggested fundamentally different patterns of acute demyelination across patients [7] and a dynamic continuum of lesion activity, from acute inflammation through chronic active and inactive stages, each with distinct implications for disease progression [8, 9, 10]. Chronic active lesions, characterized by a rim of iron‐laden activated microglia surrounding a demyelinated core, are a hallmark of smoldering neuroinflammation and predict disability accumulation [9, 11, 12]. Pathology also extends well beyond visible lesions: normal‐appearing white matter (NAWM) harbors diffuse axonal injury, microglial activation, and subtle demyelination detectable even in early disease [13, 14, 15]. Conventional MRI is exquisitely sensitive for detecting MS lesions but cannot resolve the distinct pathological processes that coexist within a single voxel. Susceptibility‐based approaches (e.g., paramagnetic rim lesions) and slowly expanding lesion mapping have improved in vivo identification of chronic active lesions [9, 16], but they classify a lesion as rim‐positive or rim‐negative. This binary readout does not quantify the cellular, fiber, and extracellular compartments within the lesion or track the graded microstructural change that accompanies treatment [17], whereas diffusion basis spectrum imaging (DBSI) resolves these compartments continuously. Multi‐compartment biophysical diffusion models applied to multi‐shell diffusion data (e.g., neurite‐orientation dispersion and density imaging [18], free‐water imaging [19], etc.) have enhanced microstructural sensitivity and specificity, but most assume a fixed number of tissue compartments and do not explicitly model the inflammatory cellularity that defines active MS pathology [20, 21, 22]. DBSI was developed to address these limitations by decomposing the diffusion signal into a continuous spectrum of anisotropic (fiber‐like) and isotropic (cellular and edematous) compartments within each voxel [20]. Rather than using a priori assumptions about compartment number, DBSI estimates the voxel‐wise fraction of signal arising from coherent fibers (fiber ratio), restricted isotropic diffusion from cellular elements (cell ratio), and hindered isotropic diffusion reflecting extracellular space (hindered ratio). DBSI‐derived metrics have been histopathologically validated in experimental autoimmune encephalomyelitis (EAE) [23], cuprizone‐treated mice [20], and human MS tissue [21, 24], and can differentiate MS subtypes [25] and classify lesion types [26]. By simultaneously resolving fiber integrity, cellularity, and edema within and across lesions, DBSI offers a framework to understand how tissue microstructure responds to early disease‐modifying treatment and whether lesion‐intrinsic heterogeneity maps onto the clinical spectrum. Here, we used DBSI to characterize the microstructure of white matter lesions, NAWM, and normal‐appearing gray matter (NAGM) in treatment‐naïve patients with early relapsing MS (RMS), patients with established non‐relapsing MS (NRMS), and healthy controls (HC). We evaluated short‐term disease‐modifying therapy (DMT) effects on tissue microstructure, tested whether whole‐lesion DBSI metrics were associated with clinical disability (expanded disability status scale [EDSS]). And because whole‐lesion averages may obscure clinically relevant signal, we examined core‐rim‐perilesional topographic gradients and applied unsupervised clustering to identify biologically coherent lesion subtypes and their associations with disease course and disability. Materials and Methods Participants and Regulatory Approval The study cohort has been described previously [27]; key eligibility criteria are summarized here. HC were recruited from the community and were free of serious neurological disease. Patients with MS were recruited from the John L. Trotter MS Center at Washington University in St. Louis and carried diagnoses confirmed by treating neurologists per the 2017 revised McDonald Criteria [28]. RMS patients were enrolled at their diagnostic visit and imaged within 2 months of diagnosis and no more than one year from their initial clinical presentation; all were treatment‐naïve except for steroids (last dose ≥ 30 days before imaging) and were enrolled in anticipation of initiating either B‐cell‐depleting or sphingosine‐1‐phosphate receptor (S1PR)‐modulating therapy. NRMS patients had disease duration exceeding 10 years with no clinical relapse and no radiographic relapse (defined as no new or enlarging T2/FLAIR lesions and no gadolinium‐enhancing lesions on clinical surveillance MRI) in the preceding 3 years and were enrolled irrespective of treatment status or evidence of progression. Neurological disability was assessed using the EDSS. The Human Research Protection Office approved the study, and all participants provided written informed consent. Image Acquisition and Alignment All MRI data were acquired on the same 3‐Tesla Siemens Biograph mMR using T1‐weighted MPRAGE, T2‐weighted FLAIR, and multi‐shell diffusion sequences optimized for DBSI modeling (acquisition parameters in Table S1) [20]. MPRAGE images were processed using FreeSurfer (v7.3) [29] to generate cortical and subcortical parcellations and tissue segmentations, then registered to native diffusion space using boundary‐based registration. White matter lesions (WML; referred to as “lesions” hereafter) were segmented on co‐registered FLAIR images using TrUE‐Net [30], with manual review and correction by a trained rater. NAWM was defined as cerebral white matter voxels with no lesion overlap, after applying a 2‐mm erosion to minimize partial volume effects. NAGM was defined as cortical gray matter voxels with no visible lesion overlap. All registrations were inspected visually. Diffusion Basis Spectrum Imaging DBSI reconstructs the multi‐shell diffusion MRI signal within each voxel into a spectrum of anisotropic (fiber‐like) and isotropic (cellular and edematous) components using a constrained linear model [20]. The signal is modeled as a mixture of discrete anisotropic tensors spanning orientations and fractional anisotropies and a continuous spectrum of isotropic Gaussian components spanning low to high diffusivities, enabling simultaneous estimation of biologically distinct microstructural features within the same voxel without a priori assumptions about fiber populations. Three DBSI metrics were extracted: (1) Fiber ratio, the fraction of signal attributable to anisotropic, fiber‐like compartments, reflecting axonal and white matter tract integrity; (2) Cell ratio, the fraction attributable to restricted isotropic diffusion (low diffusivity), corresponding to cellular elements including inflammatory infiltrates and glia; and (3) Hindered ratio, the fraction attributable to moderately restricted isotropic diffusion (intermediate diffusivity), reflecting increased extracellular space from edema or tissue disruption. DBSI additionally estimates a free‐water and a fast, perfusion‐like isotropic fraction; both were negligible in lesion tissue (medians < 0.01) and are not analyzed separately. These five ratios constitute the full set of compartmental signal fractions, which sum to one within each voxel. DBSI maps were generated in native diffusion space and sampled within tissue masks (NAGM, NAWM, lesion). For topographic analysis, each lesion was partitioned by the Euclidean distance of its voxels to the lesion boundary: the outer 50% of voxels by depth formed the rim and the deeper 50% the core, with lesions smaller than 5 voxels (and thin lesions lacking interior voxels) assigned rim only. Perilesional white matter was defined as cerebral white‐matter voxels within 0–2 mm of the lesion boundary. Core, rim, and perilesional DBSI values were available for 82%, 100%, and 96% of baseline lesions, respectively. Statistical Analysis Statistical analyses were performed in R (v4.3) [31]. Group differences in demographic and clinical characteristics were assessed using one‐way ANOVA for continuous normally distributed variables, Kruskal‐Wallis or Wilcoxon rank‐sum tests for non‐normally distributed continuous variables, Fisher's exact test for categorical variables, and chi‐squared tests for count distributions. Cross‐sectional group differences in DBSI metrics across tissue compartments (NAGM, NAWM, and lesions) were evaluated using ANCOVA with age and sex as covariates. Pairwise contrasts were HC versus treatment‐naïve RMS and HC versus NRMS in NAGM and NAWM; for lesions, RMS versus NRMS was additionally evaluated. Short‐term treatment effects in RMS were assessed using paired t‐tests on each DBSI metric and tissue region; only subjects with data at both timepoints were included. Within‐subject differences between NAWM and lesions were evaluated using linear mixed‐effects models (LME) with fixed effects for tissue type (NAWM vs. lesions), clinical condition (RMS pre‐tx vs. RMS post‐tx vs. NRMS), and their interaction, and a random intercept per subject. Post hoc contrasts comparing lesions versus NAWM within each condition were estimated using the emmeans package [32]. Associations between lesional DBSI metrics and EDSS were examined using multivariable linear regression. For cross‐sectional analyses, lesional microstructure was referenced against the mean cerebral white matter of HC, adjusted for age, sex, total lesion volume, and diagnosis group (RMS vs. NRMS), including a group‐by‐EDSS interaction term. For longitudinal analyses in paired RMS subjects, lesion microstructural change between sessions was related to long‐term EDSS change, adjusting for age, sex, follow‐up time, and baseline lesion volume. Lesion topography analyses used lesion‐level LME models with random intercepts for lesions nested within subjects. The cross‐sectional model included fixed effects for group, topographic compartment (core, rim, and perilesional white matter), and their interaction; the longitudinal model substituted session (pre‐ versus post‐treatment) for group. Omnibus fixed effects were evaluated using Type III ANOVA with Satterthwaite degrees‐of‐freedom approximation. When the group‐by‐topography or time‐by‐topography interaction was significant, post hoc pairwise contrasts within each compartment were Holm‐corrected for multiple comparisons. For unsupervised lesion clustering, the three lesion‐level DBSI ratios (fiber, cell, and hindered), each computed as a single per‐lesion mean, were clustered with a Gaussian mixture model (GMM; using the mclust package [33]). The number of mixture components and the covariance parameterization were selected automatically by the Bayesian information criterion (BIC), evaluated over one‐ to nine‐component models and mclust's covariance parameterizations (Figure S5); BIC was maximized for the four‐component solution. The model was trained on baseline RMS and NRMS lesions and applied to all lesions by posterior classification. Principal component analysis (PCA) was used only to visualize the clusters in two dimensions (Figure 5A,B). Cluster microstructural profiles were summarized by z‐scoring DBSI ratios across the pooled baseline lesion population and averaging within each cluster. Spatial distributions were visualized by nonlinearly warping subject‐level cluster masks to Montreal Neurological Institute (MNI) space [34] to generate voxel‐wise frequency maps. Lesion‐level transitions between pre‐ and post‐treatment sessions were tracked using lesion identifiers. Within‐subject longitudinal changes in cluster composition were assessed via permutation testing (5000 permutations) with subject‐level label flipping; a volume‐weighted sensitivity analysis is shown in Figure S4. Between‐group differences in the proportion of total lesion volume per cluster were evaluated with Wilcoxon rank‐sum tests. Unsupervised clustering reveals distinct lesion microstructural subtypes with differential spatial and temporal characteristics. (A) PCA projection of baseline lesion DBSI metrics (fiber, cell, hindered ratios) with GMM‐assigned cluster labels (4 clusters selected by BIC). (B) Marginal density of the trained GMM along PC1 and PC2. (C) Subject‐frequency maps in MNI space: color intensity gives the proportion of subjects with a lesion of each cluster at each voxel (per‐cluster color scale shown), displayed as sagittal, coronal, and axial views at a common cut (x = 0, y = −18, z = 16). (D) Cluster microstructural fingerprints (z‐scored DBSI metrics). (E) Lesion‐level cluster transitions following treatment in paired. (F) Within‐subject changes in cluster composition (permutation p‐values annotated). (G) Baseline cluster composition in RMS vs. NRMS (Wilcoxon rank‐sum p‐values annotated). BIC, Bayesian information criterion; DBSI, diffusion basis spectrum imaging; MNI, Montreal Neurological Institute; NRMS, non‐relapsing multiple sclerosis; PC, principal component; PCA, principal component analysis; Post‐Tx, approximately 3 months after treatment initiation; Pre‐Tx, before treatment initiation; RMS, relapsing multiple sclerosis. Results Study Cohort Thirty‐seven participants were enrolled: 11 healthy controls (HC), 15 with treatment‐naïve relapsing MS (RMS), and 11 with non‐relapsing MS (NRMS). Of the 15 treatment‐naïve RMS, 13 completed paired post‐treatment imaging; 17 RMS scans contributed to the post‐treatment overall. Groups differed in age (mean ± SD: HC 28.2 ± 5.7, RMS 39.4 ± 8.8, NRMS 47.3 ± 6.5 years; p < 0.001) but not in sex distribution (HC 91%, RMS 80%, NRMS 73% female; p = 0.681). Among MS participants, NRMS had greater neurological disability and disease duration with little overlap with RMS (EDSS median [IQR]: RMS 1.5 [1.0–2.0] vs. NRMS 4.0 [2.5–6.5]; p = 0.007) and longer disease duration (RMS 0.4 [0.3–2.1] vs. NRMS 14.1 [11.6–20.4] years; p = 0.002), reflecting the intended clinical contrast between early relapsing and established non‐relapsing disease (Table 1). Across the 26 MS participants, 726 WMLs were analyzed at baseline; lesion counts and sizes were comparable between groups (Table S3). TABLE 1Demographic characteristics of the study cohort at baseline.HC (n = 11)RMS (n = 15)NRMS (n = 11) p Age (years), mean ± SD28.2 ± 5.739.4 ± 8.847.3 ± 6.5< 0.001 † Female, n (%)10 (91%)12 (80%)8 (73%)0.681 ‡ EDSS, median [IQR]—1.5 [1.0, 2.0]4.0 [2.5, 6.5]0.007 § Disease duration (years), median [IQR]—0.4 [0.3, 2.1]14.1 [11.6, 20.4]0.002 § Abbreviations: EDSS, expanded disability status scale; IQR, interquartile range; SD, standard deviation. † One‐way analysis of variance (ANOVA). ‡ Fisher's exact test. § Wilcoxon rank‐sum. Demographic characteristics of the study cohort at baseline. Abbreviations: EDSS, expanded disability status scale; IQR, interquartile range; SD, standard deviation. One‐way analysis of variance (ANOVA). Fisher's exact test. Wilcoxon rank‐sum. Global Tissue Microstructure: Group Differences and Short‐Term Treatment Effects DBSI‐derived fiber, cell, and hindered ratios were compared across HC, RMS, and NRMS in NAGM, NAWM, and WML. Short‐term treatment effects (~3 months following initiation of B‐cell‐depleting or S1PR‐modulating therapy) were evaluated in 13 paired RMS participants (Figure 1). DBSI detects microstructural abnormality and early treatment effects across tissue compartments. Group differences (HC, RMS, NRMS) were assessed using linear models adjusted for age and sex (black p‐values); short‐term treatment effects (~3 months) were evaluated using paired t‐tests (red p‐values). HC shown in gray, RMS in blue, NRMS in orange; red lines connect paired RMS subjects. (A–C) Fiber ratio, (D–F) cell ratio, and (G–I) hindered ratio in NAGM, NAWM, and WML, respectively. HC, healthy controls; NAGM, normal‐appearing gray matter; NAWM, normal‐appearing white matter; NRMS, non‐relapsing multiple sclerosis; Post‐Tx, post‐treatment; Pre‐Tx, pre‐treatment; RMS, relapsing multiple sclerosis; WML, white matter lesion. In NAGM, no group differences reached significance for any metric. Cell ratio was similar between HC and treatment‐naïve RMS at baseline and numerically lower in NRMS (HC vs. RMS p = 0.889; HC vs. NRMS p = 0.795) but decreased significantly in RMS following treatment (p = 0.024), suggesting that gray matter inflammatory cellularity is sensitive to early DMT. Hindered ratio in NAGM showed no group or treatment differences (Figure 1A,D,G). In NAWM, treatment‐naïve RMS exhibited lower fiber ratio compared with HC (p = 0.041), whereas NRMS did not differ from HC (p = 0.249), and no treatment‐related change in fiber ratio was observed (p = 0.258). Cell ratio in NAWM showed a trend toward elevation in both RMS (p = 0.06) and NRMS (p = 0.27) compared with HC, with no treatment effect (p = 0.255). Hindered ratio in NAWM was elevated in treatment‐naïve RMS relative to HC (p = 0.022), with no treatment effect (p = 0.329; Figure 1B,E,H). In lesions, fiber ratio increased significantly following treatment (p = 0.012), consistent with partial structural recovery or reduced demyelination; baseline RMS and NRMS did not differ (p = 0.332). Lesional cell ratio decreased significantly with treatment (p = 0.025), with comparable RMS and NRMS baselines (p = 0.262), indicating rapid reduction of inflammatory cellularity. Hindered ratio in lesions was numerically higher in NRMS than RMS at baseline (p = 0.068) and unchanged with treatment (p = 0.734; Figure 1C,F,I). Within‐Subject Microstructural Differences Between NAWM and MS Lesions Having established group‐level differences relative to HCs, we next examined microstructural differences within individual MS participants between their own lesion tissues and surrounding NAWM (Figure 2). Lesions had a markedly lower fiber ratio than NAWM in RMS pre‐treatment (N = 15), RMS post‐treatment (N = 17), and NRMS (N = 11; all p < 0.001), lower cell ratio in all three conditions (all p < 0.001), and a higher hindered ratio (all p < 0.001). The lesion microenvironment is therefore consistently distinguished from surrounding normal‐appearing tissue by reduced fiber integrity and increased extracellular space, regardless of treatment status or disease course. Notably, cell ratio was lower in lesions than NAWM, a counterintuitive finding given the expected inflammatory cellularity in MS lesions [9]. Lesional microstructure diverges from normal‐appearing white matter regardless of disease course. Tissue effects (NAWM vs. WML) were evaluated using linear mixed‐effects models with a random intercept for subject. Points and connecting lines indicate within‐subject NAWM‐WML pairs. (A) Fiber ratio, (B) cell ratio, and (C) hindered ratio across RMS pre‐treatment, RMS post‐treatment, and NRMS. NAWM, normal‐appearing white matter; NRMS, non‐relapsing multiple sclerosis; Post‐Tx, post‐treatment; Pre‐Tx, pre‐treatment; RMS, relapsing multiple sclerosis; WML, white matter lesions. Whole‐Lesion Microstructure Shows Limited Association With Clinical Disability Whole‐lesion DBSI metrics showed limited association with clinical disability (Figure 3). Cross‐sectionally, HC‐referenced lesional fiber, cell, and hindered ratios were not significantly associated with baseline EDSS in either RMS or NRMS after adjustment for age, sex, and total lesion volume (Figure 3A–C). Cell ratio in RMS showed the strongest directional trend, with lower lesional cellularity associated with greater disability (β = −0.023, p = 0.065). Similar null results were observed when lesion metrics were referenced against each participant's own NAWM (Figure S1). Longitudinally, short‐term changes in lesional DBSI metrics did not predict long‐term EDSS trajectory in paired RMS subjects (Figure 3D–F). These null associations have three possible explanations: lesion microstructure may be genuinely uncoupled from EDSS, whole‐lesion summary metrics may obscure the compartments and lesion subtypes that carry the clinical signal, or the input of lesions in the spinal cord, which likely influence EDSS disproportionately, were not assessed here. Because MS lesions are well established to be microstructurally heterogeneous within and across lesions [7, 12], we next tested whether finer‐grained spatial and lesion‐level analyses recover associations that whole‐lesion averages do not. Whole‐lesion microstructure shows limited association with clinical disability. Associations were assessed using multivariable linear models adjusted for age, sex, and total lesion volume (cross‐sectional) or follow‐up time and baseline lesion volume (longitudinal). Shaded bands represent 95% confidence intervals; adjusted β and p‐values are annotated. (A–C) Baseline lesional fiber, cell, and hindered ratios referenced to healthy control white matter versus baseline EDSS in RMS (blue) and NRMS (orange). (D–F) Short‐term change in lesional fiber, cell, and hindered ratios vs. long‐term EDSS change in paired RMS subjects. EDSS, expanded disability status scale; HC, healthy controls; NRMS, non‐relapsing multiple sclerosis; RMS, relapsing multiple sclerosis. Lesion Topography Distinguishes Disease Course and Treatment Effects To characterize spatial heterogeneity, lesional and peri‐lesional voxels were classified into three compartments: lesion core (inner 50%), lesion rim (outer 50%), and perilesional white matter (0–2 mm beyond the lesion boundary; Figure 4A,B). Lesion topography distinguishes disease course and reveals treatment‐responsive microstructural gradients. (A) Schematic of lesion topography: core (inner 50% by Euclidean distance), rim (outer 50%), and perilesional white matter (0–2 mm outside lesion boundary). (B) Representative FLAIR image with corresponding topography masks. (C–E) Cross‐sectional comparison of fiber, cell, and hindered ratios between RMS and NRMS by topographic compartment. (F–H) Longitudinal pre‐ versus post‐treatment comparison within RMS. All models used lesion‐level linear mixed‐effects with random intercepts for lesions nested within subjects. Estimated marginal means ±95% CI shown; omnibus Type III tests displayed beneath each panel. p‐values above plots in (D) represent Holm‐adjusted post hoc pairwise contrasts, applied where the group‐by‐topography interaction was significant. NRMS, non‐relapsing multiple sclerosis; Peri‐lesion, perilesional white matter (0–2 mm); Post‐Tx, approximately 3 months after treatment initiation; Pre‐Tx, before treatment initiation; RMS, relapsing multiple sclerosis. Cross‐sectionally, the main effect of topography was significant for all three metrics (all p < 0.001), confirming systematic variation across compartments. For fiber ratio, neither the main effect of group (p = 0.341) nor the group‐by‐topography interaction (p = 0.153) was significant (Figure 4C). For cell ratio, the main effect of group approached significance (p = 0.057), and the group‐by‐topography interaction was significant (p < 0.001); Holm‐adjusted contrasts showed RMS > NRMS at the lesion core (p = 0.032) and rim (p = 0.044) but not in perilesional white matter (p = 0.137; Figure 4D). Inflammatory cellularity therefore separates relapsing from non‐relapsing disease specifically within the lesion body. For hindered ratio, neither the main effect of group (p = 0.239) nor the group‐by‐topography interaction (p = 0.696) was significant (Figure 4E). Longitudinally in paired RMS, the main effect of topography was significant for all three metrics (all p < 0.001), and the main effect of time was significant for all three (all p < 0.001), with fiber ratio increasing and cell and hindered ratios decreasing after treatment. None of the time‐by‐topography interactions reached significance (fiber p = 0.971; cell p = 0.842; hindered p = 0.897; Figure 4F–H), confirming that early DMT produced similar microstructural improvement across core, rim, and perilesional white matter. Unsupervised Lesion Clustering Reveals Microstructural Subtypes With Distinct Disease and Treatment Profiles To move beyond region‐level summaries and capture lesion‐intrinsic heterogeneity, we applied PCA to DBSI‐derived fiber, cell, and hindered ratios from baseline RMS and NRMS lesions, followed by GMM clustering of the ratios (Figure 5A,B). Four distinct clusters emerged (Table 2), each cluster corresponds to a recognizable microstructural state, defined by its z‐scored fiber, cell, and hindered profile relative to the pooled baseline lesion population (Figure 5D). Cluster 1 was edematous and hypocellular (high hindered, low cell and fiber), cluster 2 near‐average across all three metrics, cluster 3 actively inflamed (high cell and hindered, depleted fiber), and cluster 4 structurally preserved (high fiber, low hindered); z‐scores for each cluster are given in Table 2 and Figure 5D. TABLE 2Lesion cluster characteristics.Cluster 1Cluster 2Cluster 3Cluster 4 p Contributing subjects, n 23262324Baseline lesion count105317105199< 0.001 † RMS, n (%)46 (44%)158 (50%)61 (58%)132 (66%)NRMS, n (%)59 (56%)159 (50%)44 (42%)67 (34%)Lesion volume (mm3), median [IQR]201.6 [61.4, 592.7]80.1 [38.7, 222.9]46.7 [25.4, 101.4]42.7 [24.0, 74.8]< 0.001 † DBSI‐derived ratiosFiber ratio, median [IQR]0.511 [0.388, 0.575]0.584 [0.513, 0.644]0.394 [0.340, 0.466]0.663 [0.607, 0.724]< 0.001 ‡ Cell ratio, median [IQR]0.101 [0.064, 0.134]0.243 [0.187, 0.287]0.301 [0.238, 0.372]0.239 [0.191, 0.279]< 0.001 ‡ Hindered ratio, median [IQR]0.234 [0.195, 0.301]0.155 [0.133, 0.179]0.258 [0.230, 0.295]0.092 [0.073, 0.111]< 0.001 ‡ Abbreviations: DBSI, diffusion basis spectrum imaging; IQR, interquartile range; NRMS, non‐relapsing multiple sclerosis; RMS, relapsing multiple sclerosis. † Chi‐squared. ‡ Kruskal‐Wallis. Lesion cluster characteristics. Abbreviations: DBSI, diffusion basis spectrum imaging; IQR, interquartile range; NRMS, non‐relapsing multiple sclerosis; RMS, relapsing multiple sclerosis. Chi‐squared. Kruskal‐Wallis. Spatial frequency maps in MNI space (Figure 5C) showed partially distinct neuroanatomical distributions. Cluster 1 (edematous, hypocellular) lesions concentrated in periventricular white matter and along the corpus callosum. Cluster 2 (near average) showed substantial spatial overlap with cluster 1, suggesting that microstructurally intermediate lesions tend to occupy similar territories as the more pathologically severe cluster 1 lesions. Clusters 3 and 4 were sparsely distributed across subjects, indicating that actively inflamed and structurally preserved subtypes are comparatively rare in this cohort. Lesion‐level cluster transitions following treatment in paired RMS subjects revealed that a substantial proportion of lesions remained in their baseline cluster while a subset transitioned to clusters with higher fiber and lower cell ratios, consistent with microstructural improvement (Figure 5E and Table S2). Within‐subject proportional representation of lesions across the four clusters did not shift significantly with treatment (Figure 5F), though directional trends were observed amid considerable between‐subject variability; a volume‐weighted sensitivity analysis yielded similar results (Figure S4). At baseline, RMS and NRMS showed non‐significant but directionally distinct distributions of total lesion volume across clusters (Figure 5G): the edematous, hypocellular cluster 1 phenotype represented a greater proportion of lesion volume in NRMS than RMS, while the actively inflamed cluster 3 phenotype was proportionally more prevalent in RMS than NRMS, suggesting that lesion microstructural phenotypes may differ across disease course. Discussion This study used DBSI to characterize WML, NAWM, and NAGM in two contrasting MS clinical cohorts (treatment‐naïve early relapsing disease versus established non‐relapsing disease). By reconstructing the diffusion signal into fiber, cell, and hindered compartments within a single voxel [20, 21], DBSI provides a multi‐dimensional assessment of tissue pathology that extends beyond conventional MRI and diffusion tensor imaging. Four principal findings emerge: (1) diffuse microstructural injury is present in NAWM at the earliest clinically detectable stage of RMS; (2) DMT rapidly alters cellularity in both lesions and gray matter while increasing the lesional fiber signal; (3) inflammatory cellularity distinguishes relapsing from non‐relapsing disease primarily within the lesion interior; and (4) unsupervised clustering identifies biologically interpretable lesion subtypes whose distributions differ across disease course. Diffuse White Matter Injury Is Detectable at the Earliest Clinical Stage Treatment‐naïve RMS exhibited significantly reduced fiber ratio and elevated hindered ratio in NAWM relative to HC, consistent with growing evidence that NAWM in MS harbors diffuse microstructural abnormalities detectable at the earliest clinical stages [14, 15]. Histopathology has confirmed that NAWM in MS contains chronic axonal injury, microglial activation, and subtle demyelination behind an apparently intact blood–brain barrier [13, 14]. The fiber ratio reduction likely reflects subclinical demyelination [35], axonal loss [36], and disruption of fiber coherence [37], while the elevated hindered ratio may correspond to extracellular matrix expansion in NAWM [20, 21, 35]. Previous DBSI work reported a gradation of increasing edema, demyelination, and axonal loss from clinically isolated syndrome through relapsing–remitting to progressive MS [38]. NAWM differences between NRMS and HC did not reach statistical significance, despite NRMS patients' longer disease duration and greater disability. This likely reflects the older age of the NRMS group, limited statistical power, and the fact that most NRMS were on DMT, which may have attenuated diffuse white matter pathology relative to the treatment‐naïve RMS group. The absence of significant NAGM differences across groups suggests cortical microstructural changes may require larger samples, higher spatial resolution, or more sensitive metrics to detect at the tissue‐averaged level assessed here. DMT Rapidly Alters Lesional Cellularity and Fiber Integrity Treatment effects observed ~3 months after initiating B‐cell‐depleting or S1P receptor‐modulating therapy demonstrate that DBSI is sensitive to early changes in lesional microstructure: lesional fiber ratio increased, cell ratio decreased, and cortical NAGM cell ratio also declined. Immunomodulatory therapy thus rapidly modulates the cellular inflammatory milieu within lesions and gray matter, with partial restoration of the anisotropic signal compartment in lesions. The lesional fiber ratio increase could reflect reduced inflammatory edema unmasking intact fiber structure, early remyelination, or both. Because DBSI ratios are voxel‐wise signal fractions that sum to one, they are compositional and interdependent: a treatment‐related decrease in the cell and hindered fractions necessarily raises the fiber fraction. In a within‐lesion decomposition of the paired data (Table S4), the mean lesional fiber ratio increase was fully accounted for by concurrent decreases in the isotropic compartments with no residual gain beyond redistribution. The fiber ratio increase should therefore be read primarily as reduced inflammatory cellularity and edema unmasking preserved fiber signal. Although these compositional metrics cannot establish remyelination, myelin turnover on a comparable timescale is detectable in vivo [39], so an early remyelination contribution remains biologically plausible but cannot be confirmed here. The cell ratio decrease is consistent with reduced density of inflammatory infiltrates, including lymphocytes, activated macrophages, and microglia, following effective immunosuppression [20, 23]. These findings are concordant with prior EAE studies showing that immunomodulation selectively reduces the restricted isotropic compartment while preserving or restoring anisotropic components [23]. NAWM fiber and hindered ratios did not change significantly over this short interval, suggesting diffuse white matter pathology may be slower to reverse than focal lesional inflammation, potentially reflecting structural damage (axonal loss, gliosis) not readily amenable to short‐term immunomodulation. This finding is consistent with the emerging understanding that smoldering inflammation and neurodegeneration in MS are driven by compartmentalized innate immune processes that current peripheral immunotherapies may incompletely address [9, 17]. Anti‐CD20 agents, for example, suppress new inflammatory activity but do not resolve chronic active lesions over multi‐year follow‐up [17], highlighting the gap between acute inflammatory suppression and chronic tissue repair. The limited whole‐lesion association with disability is consistent with the clinico‐radiological paradox in MS [40]. EDSS coarseness [41, 42] and the likelihood that principal disability drivers lie beyond brain lesions [14, 38] constrain the variance that lesion‐intrinsic metrics can explain. The directional trend for lesional cell ratio is nonetheless biologically plausible [9, 16] and merits replication in larger cohorts with extended follow‐up and disability endpoints beyond EDSS. Cellularity Distinguishes Disease Stage Within the Lesion Interior Topographic analyses add spatial granularity to the group‐level findings: cell ratio was higher at the lesion core and rim in RMS than NRMS, while perilesional cell ratio did not differ, consistent with acute inflammatory cellularity concentrated within the lesion body in relapsing disease that diminishes as lesions become chronic. This is concordant with chronic active lesions (CALs), where a hypocellular center with reduced axonal fibers is surrounded by a rim of activated microglia and macrophages [9, 43]. These topographic gradients also clarify why the whole‐lesion cell ratio fell below NAWM (Figure 2B), which reflects two compatible factors: cellularity is concentrated at the rim and diluted by a hypocellular, demyelinated core, and the restricted‐isotropic (cell) signal indexes total restricted cellular volume, such as reduced in axon‐ and glia‐depleted chronic lesion tissue, rather than inflammatory cells specifically. That the cell ratio nonetheless rises with active disease (higher in the RMS lesion core and rim than NRMS) and falls with treatment indicates the metric tracks inflammatory cellularity where it is present. Elevated core and rim cellularity in RMS, with unchanged perilesional white matter, suggests cellular content of the lesion interior declines as lesions transition from active to chronic inactive states. The fiber and hindered ratios showed similar topographic profiles in RMS and NRMS, suggesting that structural damage and edema, once established, are relatively uniform across lesion compartments regardless of disease stage. In contrast, cellularity is the dimension along which disease course and, by extension, the predominant immune process, are most clearly distinguished. This dissociation between inflammatory and structural dimensions of lesion heterogeneity underscores the value of DBSI's multi‐compartment decomposition, which can parse these pathologically distinct processes within the same voxel [20, 21]. Treatment improved fiber, cell, and hindered ratios comparably across core, rim, and perilesional white matter rather than targeting any single compartment. This contrasts with the CAL model, in which rim‐localized inflammation is relatively resistant to peripheral immunotherapy [9, 17]. The discrepancy likely reflects timing: our RMS cohort was imaged shortly after treatment initiation, before compartmentalized, treatment‐refractory innate immune processes are thought to establish at the lesion edge. Paramagnetic rims are consistently present in recently formed lesions and may gradually diminish as lesions age [16, 44], suggesting treatment‐naïve RMS lesions represent a predominantly acute inflammatory phenotype more responsive to systemic immunotherapy than the entrenched smoldering inflammation of established CALs. Unsupervised Clustering Identifies Biologically Coherent Lesion Phenotypes The unsupervised GMM clustering provides a complementary, data‐driven approach to lesion characterization. The four identified clusters: relatively characterized as edematous‐hypocellular (cluster 1), microstructurally average (cluster 2), actively inflamed (cluster 3), and structurally preserved (cluster 4) have intuitive biological interpretations that align with known MS lesion pathology. Prior supervised DBSI classification achieved > 93% concordance on MS lesion subtypes [26]. The present approach extends this work by identifying phenotypes without predefined labels. Earlier work has similarly shown that fiber and cell ratios are the most important DBSI metrics for classifying MS clinical subtypes [25], reinforcing the biological relevance of the fiber integrity and cellularity dimensions that drive our cluster definitions. Cluster 1 (edematous‐hypocellular) lesions concentrated in periventricular white matter and along the corpus callosum, regions with known predilection for MS lesions due to proximity to cerebrospinal fluid‐borne inflammatory mediators and the periventricular venous architecture [13]. That cluster 3 (actively inflamed) was proportionally more prevalent in RMS and cluster 1 more prevalent in NRMS, while not statistically significant, is directionally consistent with lesion microstructure evolving from an actively inflammatory state in early relapsing disease toward a more quiescent, edematous state in established non‐relapsing disease [38, 45]. Emerging evidence has identified a novel “broad rim” lesion type, marked by extensive innate immune activation, that is associated with rapid disease progression [43], further underscoring the importance of characterizing lesion heterogeneity with tools that resolve inflammation, fiber integrity, and edema. DBSI metrics here were averaged across the entire lesion, which may dilute the spatial gradients relevant to established pathological classifications. CALs, defined by a rim of activated microglia surrounding a demyelinated core [9, 43], would have their core‐rim gradient collapsed into a single intermediate value, potentially sorting such lesions into the microstructurally intermediate cluster 2 rather than into clusters reflecting their rim‐dominant pathology. The topographic analyses (Figure 4) partially address this by resolving core, rim, and perilesional compartments, but the clustering itself operates on spatially averaged metrics. Future work incorporating voxel‐wise or shell‐wise clustering may better resolve this heterogeneity [46]. Several caveats apply to interpreting DBSI. First, as a computational model, its outputs depend on modeling choices: the sampled isotropic diffusivity spectrum and the diffusivity ranges used to summarize the restricted (cell), hindered, and free‐water compartments. Consequently, absolute DBSI metric values should not be assumed to be directly interchangeable across studies that use different acquisition protocols, preprocessing pipelines, fitting algorithms, isotropic‐diffusivity ranges, or compartment definitions. Second, we also clarify that DBSI‐derived fractions represent signal fractions estimated by a multicomponent model. Because these fractions are compositional and constrained by the total model signal, changes in one fraction may be accompanied by changes in other fractions. Accordingly, individual DBSI metrics should be interpreted in the context of the full metric profile rather than as fully independent biological measures. Third, DBSI metrics are histopathologically validated surrogates of demyelination, axonal integrity, and cellularity, but the restricted‐isotropic (cell) compartment reflects the aggregate signal contribution from restricted isotropic water environments [47] and should not be interpreted as identifying a specific type of cellular population without appropriate histopathologic or molecular validation. Finally, the reproducibility of DBSI metrics in MS tissue is still being established, and our longitudinal changes should be interpreted within that developing evidence base. Limitations The 37‐participant cohort was intentionally enriched for clinical contrast between early relapsing and established non‐relapsing disease, yielding robust effect sizes for primary comparisons but limited power for interaction analyses and clinical associations; replication in larger, multi‐site cohorts is warranted. Cross‐sectional contrasts inform disease‐stage associations rather than causal trajectories, motivating prospective follow‐up of patients transitioning from relapsing to progressive disease. The ~3‐month treatment interval captures acute immunomodulatory response; longer follow‐up is needed to determine sustainment and predictive value for long‐term outcomes. Inclusion of both drug classes broadens generalizability but precludes resolution of drug‐class‐specific effects. Although age differed across groups by design and models adjusted for age and sex, residual confounding from non‐linear age‐microstructure relationships cannot be excluded. DBSI metrics are histopathologically validated surrogate markers of CNS white matter pathology [20, 21, 23], yet the precise cellular identity captured by the restricted isotropic component (e.g., T cells, B cells, microglia, other glia) cannot be resolved without paired histopathological data, an important avenue for future translational work. The unsupervised lesion clustering, while yielding biologically coherent subtypes, should be considered hypothesis‐generating and will benefit from validation in independent, larger datasets. Together, this study demonstrates that DBSI provides a sensitive, multi‐dimensional framework for characterizing heterogeneous MS tissue microstructure across disease stages and treatment. By resolving fiber, cellular, and extracellular compartments within and across lesions, DBSI offers a path to non‐invasive lesion phenotyping with potential applications in clinical trial design and precision MS care. Author Contributions A.K., M.R.B., Q.W., and T.L.S.B. contributed to the conceptualization and design of the study. A.K., A.L., B.J., N.S., J.R., Z.T., Y.W., Q.W., and M.R.B. contributed to the acquisition and analysis of the data. A.K., K.F., Q.W., and M.R.B. contributed to manuscript drafting and figure preparation. Funding This work was supported by the National Institutes of Health (K23 NS128325, R01 AG074909, R01 NS103988) and the National Multiple Sclerosis Society (RG‐2406‐43174). The Brier Lab is supported by generous philanthropic donations from The Laura Fund. Conflicts of Interest T.L.S.B. has received consulting fees from Biogen, Bristol Myers Squibb, Eisai, Eli Lilly, Johnson & Johnson, Merck, and Roche, and holds issued US patent 16/097,457 (Diffusion Basis Spectrum Imaging). M.R.B. has received consulting honoraria from Biogen, Bristol Myers Squibb, Roche, TG Therapeutics, Novartis, Amgen, Sanofi, and Genentech unrelated to this present work. Washington University School of Medicine has received research support from Siemens. All other authors report no potential conflicts of interest related to the present study. Supporting information Figure S1: Within‐subject referenced lesional microstructure in relation to baseline EDSS. This figure complements Figure 3A–C, which uses healthy control white matter as the reference. Figure S2: Baseline lesion cluster composition in relation to EDSS. This figure extends the cross‐sectional clinical association analyses in Figure 3A–C to cluster‐level lesion composition (Figure 5 and Table 2). Figure S3: Longitudinal change in lesion cluster composition in relation to long‐term clinical outcome. This figure extends the longitudinal clinical association analyses in Figure 3D–F to cluster‐level lesion composition. Figure S4: Sensitivity analysis: volume‐weighted within‐subject changes in cluster composition following treatment. This figure is a sensitivity analysis for the count‐based cluster composition change shown in Figure 5F. Figure S5: Bayesian Information Criterion for Gaussian mixture model selection. This figure supports the unsupervised clustering methodology used in Figure 5A and Table 2. Table S1: MRI acquisition parameters. Table S2: Lesion‐level cluster transition matrix (counts) following treatment initiation in paired RMS subjects. Table S3: Lesion burdens distribution. Table S4: Within‐lesion decomposition of the matched pre‐ and post‐treatment. Acknowledgments We thank Anne H. Cross for helpful comments on the manuscript. Funding Statement This work was funded by National Institutes of Health grants K23 NS128325, R01 AG074909, and R01 NS103988; National Multiple Sclerosis Society grant RG‐2406‐43174; The Laura Fund . Contributor Information Qing Wang, Email: wangqing@wustl.edu. Matthew R. Brier, Email: brierm@wustl.edu. Data Availability Statement De‐identified data are available from the corresponding authors upon reasonable request from qualified investigators, subject to a data use agreement covering protected health information. The DBSI analysis was performed using proprietary in‐house software developed by our group; therefore, the full codebase is not currently available as an open‐source package. Although the software is not publicly released, the DBSI analysis code and supporting scripts may be shared with qualified academic investigators through a research collaboration, subject to institutional requirements and an appropriate data/code‐use agreement. The DBSI framework and model parameters are previously described [20, 47]. References Associated Data Supplementary Materials Figure S1: Within‐subject referenced lesional microstructure in relation to baseline EDSS. This figure complements Figure 3A–C, which uses healthy control white matter as the reference. Figure S2: Baseline lesion cluster composition in relation to EDSS. This figure extends the cross‐sectional clinical association analyses in Figure 3A–C to cluster‐level lesion composition (Figure 5 and Table 2). Figure S3: Longitudinal change in lesion cluster composition in relation to long‐term clinical outcome. This figure extends the longitudinal clinical association analyses in Figure 3D–F to cluster‐level lesion composition. Figure S4: Sensitivity analysis: volume‐weighted within‐subject changes in cluster composition following treatment. This figure is a sensitivity analysis for the count‐based cluster composition change shown in Figure 5F. Figure S5: Bayesian Information Criterion for Gaussian mixture model selection. This figure supports the unsupervised clustering methodology used in Figure 5A and Table 2. Table S1: MRI acquisition parameters. Table S2: Lesion‐level cluster transition matrix (counts) following treatment initiation in paired RMS subjects. Table S3: Lesion burdens distribution. Table S4: Within‐lesion decomposition of the matched pre‐ and post‐treatment. Data Availability Statement De‐identified data are available from the corresponding authors upon reasonable request from qualified investigators, subject to a data use agreement covering protected health information. The DBSI analysis was performed using proprietary in‐house software developed by our group; therefore, the full codebase is not currently available as an open‐source package. Although the software is not publicly released, the DBSI analysis code and supporting scripts may be shared with qualified academic investigators through a research collaboration, subject to institutional requirements and an appropriate data/code‐use agreement. The DBSI framework and model parameters are previously described [20, 47].