Association of Gray Matter Atrophy Patterns With Clinical Phenotype and Progression in Multiple Sclerosis

Objectives Gay matter (GM) involvement is clinically relevant in multiple sclerosis (MS). Using source-based morphometry (SBM), we characterized GM atrophy and its 1-year evolution across different MS phenotypes. Methods Clinical and MRI data were obtained at 8 European sites from 170 healthy controls (HCs) and 398 patients with MS (34 with clinically isolated syndrome [CIS], 226 with relapsing-remitting MS [RRMS], 95 with secondary progressive MS [SPMS], and 43 with primary progressive MS [PPMS]). Fifty-seven HCs and 144 with MS underwent 1-year follow-up. Baseline GM loss, atrophy progression, and correlations with disability and 1-year clinical worsening were assessed. Results SBM identified 26 cerebellar, subcortical, sensory, motor, and cognitive GM components. GM atrophy was found in patients with MS vs HCs in almost all components (p range <0.001–0.04). Compared to HCs, patients with CIS showed circumscribed subcortical, cerebellar, temporal, and salience GM atrophy, while patients with RRMS exhibited widespread GM atrophy. Cerebellar, subcortical, sensorimotor, salience, and frontoparietal GM atrophy was found in patients with PPMS vs HCs and in patients with SPMS vs those with RRMS. At 1 year, 21 (15%) patients had clinically worsened. GM atrophy progressed in MS in subcortical, cerebellar, sensorimotor, and fronto-temporo-parietal components. Baseline higher disability was associated (R2 = 0.65) with baseline lower normalized brain volume (β = −0.13, p = 0.001), greater sensorimotor GM atrophy (β = −0.12, p = 0.002), and longer disease duration (β = 0.09, p = 0.04). Baseline normalized GM volume (odds ratio 0.98, p = 0.008) and cerebellar GM atrophy (odds ratio 0.40, p = 0.01) independently predicted clinical worsening (area under the curve 0.83). Conclusion GM atrophy differed across disease phenotypes and progressed at 1 year in MS. In addition to global atrophy measures, sensorimotor and cerebellar GM atrophy explained baseline disability and clinical worsening.

In the last decades, gray matter (GM) involvement has been increasingly recognized as a crucial component of pathophysiology in multiple sclerosis (MS). 1 GM atrophy occurs from the earliest MS phases 2 and progresses over time, with rates being 8 times greater than controls in patients with relapsing-remitting MS (RRMS) and up to 14 times greater in patients with progressive MS. 3 A substantial relationship between GM atrophy and clinical disability has been consistently demonstrated. 1 The regional distribution of GM atrophy is not homogeneous across phenotypes, with an earlier involvement of deep GM and the parietal lobe 1,4,5 and a progressive spreading of GM loss to frontal, temporal, occipital, and cerebellar regions at later disease stages. 1,6,7 One of the most widely used techniques to investigate GM atrophy localization is voxel-based morphometry, 8,9 a mass univariate method that does not consider information about the relationships among voxels other than those in the immediate neighborhood. In contrast, independent component (IC) analysis (ICA) is able to extract spatially independent sources of GM changes from a set of MRIs. 10 In particular, source-based morphometry (SBM) 11 is an ICA-based technique that allows GM maps to be decomposed into distinct patterns of GM density that covary across individuals. These patterns can be thought of as GM networks, and SBM can be used to quantify betweengroup differences of GM atrophy regarding such patterns. This is achieved by comparing participant-wise loadings (or weights), which represent the degree to which a pattern is present at an individual participant level. Reduced loadings are usually considered a measure of atrophy. 11 Studies applying SBM in MS found that GM atrophy occurs largely in a nonrandom manner 12,13 and develops in distinct anatomic patterns that show association with clinical disability 12 and, to lesser extent, with white matter (WM) damage. 13 A 10-year longitudinal SBM study in a cohort of patients with RRMS retrospectively extracted from a clinical trial 14 showed that atrophy progression in motor and cognitive GM networks was higher in clinically worsened compared to stable patients with MS. Despite these encouraging results, SBM has never been applied to large cohorts of patients with MS including all main clinical phenotypes and follow-up evaluations. We hypothesize that the use of this technique in such a population would be useful to characterize patterns of atrophy development in the main large-scale GM networks across different disease phenotypes and their relationship with clinical disability and deterioration.
In this study, we applied SBM to a large, multicenter MS dataset to determine the main patterns of GM atrophy and their 1-year evolution in patients with MS compared to healthy controls (HCs) and within the MS population according to clinical phenotype. We also evaluated the correlation between the occurrence of baseline GM atrophy in specific networks and concomitant clinical impairment, and we assessed the predictive role of GM network atrophy on disability worsening.

Standard Protocol Approvals, Registrations, and Patient Consents
Approval was received from the local ethics committees at each participating center; study participants signed an informed consent before enrollment.

Participants
We recruited study participants at 8 European sites from the Magnetic Resonance Imaging in Multiple Sclerosis (MAG-NIMS) consortium (magnims.eu): (1) the Amsterdam MS Center, Amsterdam UMC, location VUmc, Amsterdam (the Netherlands); (2) the Cemcat and Section of Neuroradiology, University Hospital Vall d'Hebron, Barcelona (Spain); (3) St. Josef Hospital Ruhr University, Bochum (Germany); (4) Institute of Neurology, UCL, London (UK); (5) the Neurocenter of Southern Switzerland, Lugano (Switzerland); (6) the Neuroimaging Research Unit, San Raffaele Scientific Institute, Milan (Italy); (7) the MRI Center "SUN-FISM," University of Campania "Luigi Vanvitelli," Naples (Italy); and (8) the Nuffield Department of Clinical Neurosciences, Oxford (UK). Participants were part of a previous study aimed at characterizing the distribution and regional evolution of cervical cord atrophy. 15 To be included, patients with MS had to have stable treatment in the past 6 months and received no corticosteroids during the last month. Patients with a clinically isolated syndrome (CIS) suggestive of MS had to have a first neurologic episode suggestive of demyelination and a clinical assessment within 3 months from symptoms. Patients were excluded if they had history of brain trauma, important comorbid conditions, drug or alcohol abuse, or any other medical conditions interfering with MRI or if they were unable to undergo MRI (claustrophobia, metal implants, pacemakers, etc.), breastfeeding, or pregnant.
Participants' evaluation included an MRI and clinical assessment at baseline and, whenever possible, after 1 year. 15 To exclude increasing disability resulting from relapses close to baseline and follow-up visits as well as consequent spurious effects on GM volumes, all patients were free from relapses and steroid treatment for at least 1 month before clinical and MRI evaluations.

Clinical Assessment
Within 2 days from MRI scanning, patients with MS underwent a neurologic evaluation by an experienced neurologist blinded to the MRI results. 15 The Expanded Disability Status Scale (EDSS) score 16 was rated, and disease-modifying treatments (DMTs) were recorded. At the 1-year follow-up, we considered a patient clinically worsened if his/her EDSS score increased ≥1.0 point when baseline EDSS score was <6.0 or if his/her EDSS score increased ≥0.5 point when baseline EDSS score was ≥6.0. 17 EDSS changes were confirmed during a second visit after 3 months. If patients did not have confirmed disability progression, they were considered clinically stable. slices with 0.9-to 1.2mm slice thickness and ≈1 mm 2 in-plane resolution) and (2) axial brain dual echo fast spin echo (TR range 2,500-4,670 milliseconds, TE range 15-27/79-120 milliseconds, FA range 90°-150°, 44 to 50 axial slices with 3-mm slice thickness and ≈0.7to 0.9-mm 2 in-plane resolution) or (3) 3-dimensional sagittal fluid-attenuated inversion recovery (TR/TE 8,000/ 125 milliseconds, TI 2,350 milliseconds, FA variable, 132 sagittal slices, thickness 1.2 mm, in-plane resolution ≈1 mm 2 ) for brain lesion assessment.

Conventional MRI Analysis
T2-hyperintense and T1-hypointense lesion volumes (LVs) of the brain were produced at baseline and follow-up with the Jim 7.0 software package (Xinapse Systems, Colchester, UK). Lesion-filled 18 T1-weighted images were used to calculate baseline normalized brain volume (NBV), normalized GM volume (NGMV), and normalized WM volume (NWMV) with FSL SIENAx. Percentage brain volume change at 1 year was calculated with FSL SIENA software. 19 Head size was measured with the inverse of the brain scaling factor derived from FSL SIENAx.

SBM Analysis
An optimized, longitudinal pipeline was implemented using a combination of voxel-based 8 and tensor-based morphometry 20 tools, as previously suggested. 14 Briefly, segmentation of baseline and 1-year (when available) 3-dimensional T1weighted images into GM, WM, and CSF was performed with SPM12. Participants having a follow-up MRI evaluation underwent an additional postprocessing step, consisting of a nonlinear registration of baseline and 1-year T1-weighted scans to generate an unbiased, participant-specific template (SPM12, pairwise registration tool). 20 We applied transformations needed to coregister baseline and follow-up images to the corresponding template to native-space GM and WM segmentations. Then, the Diffeomorphic Anatomical Registration using Exponentiated Lie algebra method was applied to rigidly aligned GM and WM images of all study participants to produce population-specific GM and WM templates and deform images to such templates. 8 Modulated images were transformed from Diffeomorphic Anatomical Registration using Exponentiated Lie to Montreal Neurological Institute space by use of an affine transformation. Finally, GM images were smoothed (8-mm full width at half-maximum gaussian kernel).
GM maps underwent SBM to produce spatial GM ICs, i.e., groups of spatially distinct GM regions showing common covariations among participants. 11 To this aim, preprocessed baseline and 1-year follow-up GM maps were all concatenated 14 and underwent spatial ICA using the SBM GIFT toolbox (trendscenter.org/trends/software/gift) and the Infomax algorithm. The GM matrix was linearly decomposed into a set of ICs, representing patterns of covarying GM volume, and a matrix of loading coefficients, representing the contribution of each scan to a component in terms of GM volume. The model order selected for SBM (i.e., n = 98) was determined from the minimum description length criterion. GM loading coefficients were extracted for subsequent statistical analysis. As previously suggested, 11,12 GM IC maps and loading coefficients of components with a main negative sign were inverted.
Statistical Analysis SAS version 9.4 (SAS Institute Inc, Cary, NC) was used for all analyses. T2 and T1 LVs were log-transformed. Comparisons of demographic, clinical, conventional MRI measures, and GM loading coefficients between HCs and patients with MS were performed with generalized linear mixed-effects models adjusted for age, sex, and head size. Such models accounted for site heterogeneity and clustering (participants within sites) using random intercepts. The same models were used for comparisons between MS phenotypes, with the following post hoc contrasts based on disease clinical evolution: HCs vs CIS, HCs vs primary progressive MS (PPMS), CIS vs PPMS, CIS vs RRMS, RRMS vs secondary progressive MS (SPMS), and PPMS vs SPMS. Post hoc contrasts were corrected for multiple comparisons using the false discovery rate method. 21 Changes over time of clinical and MRI variables were compared between HCs and patients with MS and among phenotypes using generalized linear mixed-effect models adjusted for age, sex, head size, and duration of follow-up. Site heterogeneity and clustering were accounted for with random intercepts. In the longitudinal assessment, we grouped together PPMS and SPMS because of the relatively low number of patients with follow-up assessment. This yielded the following post hoc contrasts: HCs vs CIS, CIS vs RRMS, and MS RRMS vs progressive MS. HCs and clinically worsened and stable patients with MS were also compared. To test the effect of DMT, baseline and longitudinal analyses were repeated including in the mixed-effect models the presence of DMT as a binary variable. The significance of the interaction term was also tested.
In patients with MS, we tested correlations of EDSS score with disease duration, conventional MRI variables, and GM loading coefficients using linear models adjusted for age, sex, site, DMT, and phenotype. Finally, an age-, sex-, site-, DMT-, and phenotype-corrected multivariable linear model with a combination of forward and backward stepwise variable selection identified measures independently predicting the EDSS score. The R 2 index expressed the proportion of variance explained by the model, while the proportional strength of each independent predictor was expressed by standardized coefficients (β).
Odds ratios (ORs) and related 95% confidence intervals (CIs) from logistic regression models were used to assess the association between disease progression and study variables. We adopted a Firth bias-reduced penalized-likelihood estimation approach to select the best independent predictors of disease progression using multivariate logistic regression analysis. We conducted forward and backward stepwise analyses based on penalized-likelihood ratio tests. Confounding covariates included in such regression models were follow-up duration, DMT, and center.

Data Availability
The dataset analyzed in the current study is available from the corresponding author on reasonable request.

Demographic, Clinical, and Conventional MRI Assessment
One-hundred seventy HCs and 398 patients with MS were available for the final analysis (table 1), including 34 with CIS, 226 with RRMS, 95 with SPMS, and 43 with PPMS. As expected, baseline EDSS score (p < 0.001) and disease duration (p = 0.02) were higher in patients with RRMS vs those with CIS and in patients with SPMS vs patients with RRMS (p < 0.001). Similarly, brain T2 LV and T1 LV were higher in patients with PPMS and RRMS vs patients with CIS (p range <0.001-0.006) and in those with SPMS vs those with RRMS (p = 0.006 and 0.002, respectively). Baseline NBV was lower in patients with PPMS vs HCs (p = 0.007) and in patients with Mean disease duration (range), y Median baseline EDSS score (range) In patients with MS with a follow-up evaluation, EDSS score (p = 0.5), T2 LV (p = 0.07), and T1 LV (p = 0.43) did not change over time, and percentage brain volume change did not differ between HCs and patients with MS (table 2). Clinically stable and worsened MS differed in terms of sex (p = 0.01), baseline EDSS score (p = 0.05), T2 LV (p = 0.04), and percentage brain volume change (p = 0.02), but they did not differ in terms of age (p = 0.4), disease duration (p = 0.2), and T1 LV (p = 0.06).

Baseline SBM Analysis
Among the 98 ICs estimated by SBM, 26 relevant GM components were selected by visual inspection, according to their correspondence with well-known sensorimotor and cognitive networks. 12,14 ICs provided by a previous study 22 were also used as reference templates, given the similarity of model order between that study (n = 100) and our study. Relevant GM ICs (thresholded using a z score >2.5 and a cluster extent >100 mm 3 ) were assigned to the following networks: cerebellar (4 ICs), subcortical (3 ICs      In both clinically stable and worsened patients with MS, a significant GM decrease over time was detected in the cerebellar (p = 0.02 and 0.01, respectively), sensorimotor (p = 0.002 and 0.01, respectively), and auditory (p = 0.03 and 0.02, respectively) networks, with no significant time × group or DMT × group interactions (p range 0.14-0.97). Clinically stable patients with MS showed significant GM loss in the frontoparietal network (p = 0.006).
The univariate logistic regression models showed that male sex (OR 3.2, 95% CI 1.2-11.7, p = 0.02), higher T2 (OR 2.5, After source-based morphometry with n = 98 components was run, relevant independent components (ICs) were selected and sorted into 8 subcategories: cerebellar, subcortical, sensorimotor, visual, auditory, default-mode, frontoparietal, and salience networks. Each color in the composite map corresponds to a different IC within a given subcategory. IC patterns were thresholded at z > 2.5. Images are in neurologic convention.

Discussion
Although MS is traditionally considered a WM disease, GM involvement has been recognized as a crucial determinant of clinical manifestations and prognosis. 1 Trajectories of GM atrophy development in MS are not completely understood, and a variable number of cortical and subcortical GM structures have been shown to be affected. 1 Previous studies substantially agree in showing that GM atrophy is clinically relevant, being useful in characterizing the main disease clinical phenotypes 23 and being able to explain specific disease-related symptoms, disability progression, and cognitive deficits. 1,5,7 Results of studies applying voxel-based morphometry are not fully consistent and may be limited by the small sample sizes or by the inclusion of a single disease phenotype. 6,[24][25][26][27] In this work, a rather novel multivariate method (i.e., SBM) was applied to identify spatial patterns of covarying GM volume in patients with MS. Such an analysis identified 26 distinct, nonrandom GM ICs, which ranged from patterns associated with motor or sensory related areas, to subcortical or cerebellar patterns, to regions associated with cognition (DMN, frontoparietal and salience networks). This is in line with preliminary SBM studies in MS, which were, however, limited by the lack of a follow-up examination 12,13 or by the absence of reference HCs and by the inclusion of a single disease phenotype. 14 Our components were less spatially extended than those of the previous studies, 12-14 each covering small but homogeneous brain portions. This is probably due to our relatively high ICA model order (n = 98). However, model order selection is likely to have a minor influence on results because SBM findings were consistent across a variety of different settings. 28 Compared to HCs, patients with MS exhibited significant baseline GM atrophy in almost all ICs. This is in line with previous volumetric studies 5,6,24-27 and confirms that there is a strong neurodegenerative component in MS leading to irreversible tissue loss. 1 Atrophy started to occur relatively early in the disease because patients with CIS already showed significantly GM atrophy vs HCs in subcortical and cerebellar IC and a marginal involvement of the temporal lobe. This confirms the early vulnerability to damage of subcortical GM, 4,29 which might be due to its proximity to the ventricular system-mediated pathogenetic factors. [30][31][32] Patients with RRMS exhibited a widespread pattern of GM atrophy vs HCs, involving all subcortical, cerebellar, and sensorimotor ICs, as well as some cognitive, higher-order ICs. This suggests a progressive spreading of GM pathology from subcortical to cortical areas, especially those involved in sensory and motor functions. In the same networks, diffuse GM loss became more severe in patients with SPMS vs patients with RRMS. 1,5,24 In line with previous work, patients with PPMS showed GM atrophy vs HCs mainly in sensorimotor, auditory, and subcortical networks. 26 In addition, we found GM atrophy in cerebellar, visual, frontoparietal, and salience networks. Taken together, these results suggest that SBM may be sensitive to the heterogeneity of MS-related structural damage, being powerful in detecting between-group differences of GM atrophy among different MS phenotypes. [24][25][26][27]29,33 At the 1-year follow-up, GM atrophy significantly progressed over time in patients with MS in 5 ICs, located in distinct anatomic systems. This supports the notions that GM atrophy progression in patients with MS is extended and severe and that the neurodegenerative process related to MS may lead to a premature brain aging. 34 Progression of GM atrophy was driven mainly by patients with progressive MS, who exhibited significant cerebellar, sensorimotor, frontoparietal, and temporal GM atrophy. This is in line with studies demonstrating that brain atrophy accelerates in the progressive disease phase. 3,25 On the other hand, atrophy significantly progressed over time also in patients with CIS especially in the subcortical compartment, reinforcing the notion of a preferential damage of deep GM and cerebellar structures at the initial disease stages. 4,29 Clinically worsened and stable patients with MS had a similar pattern of GM atrophy evolution. This seems to be partially counterintuitive and in contrast to previous studies, which detected a higher cortical atrophy progression in clinically worsened than in stable patients. 7,14,[35][36][37] Several factors could explain the lack of differences between stable and worsened patients, including the small number of patients with follow-up, the relatively small number of patients with clinical deterioration during the relatively short duration of the 1-year follow-up (which might not be sufficient to detect clinically relevant modifications), and the subtle contribution of relapses between baseline and follow-up to disability accumulation. Considering that brain atrophy represents an end-stage phenomenon that could develop and progress also several months after the accumulation of focal lesions due to a secondary retrograde degeneration, further studies with longer follow-up are needed to better assess possible differences between these 2 groups.
In patients with MS, a higher baseline EDSS score was significantly associated with GM atrophy of several components, belonging to all main motor, sensory, and high-order networks. This suggests that the multiparametric evaluation of atrophy of several systems is able to explain overall disease severity. 27 Moreover, the multivariate analysis found that the variables best associated with a higher EDSS score were a lower NBV, higher GM atrophy in the sensorimotor network, and longer disease duration. This indicates that GM atrophy of the sensorimotor network could be crucial to explain MS-related  clinical impairment. This is not unexpected given that the EDSS score is heavily weighted toward locomotor dysfunction.
Finally, in univariate logistic regression models, male sex, higher LVs, and lower NBV, NGMV, and GM volume in 1 cerebellar IC predicted 1-year disability worsening. The stepwise multivariable logistic model confirmed that lower NGMV and lower GM volume in the cerebellar SBM network predicted 83% of disability worsening over 1 year. These results are in line with previous studies of biomarkers for disability prediction. For instance, men with MS were shown to have a worse prognosis and a more aggressive MS course than women. 38 In other studies, baseline GM damage predicted long-term disability and cognitive deterioration at the 13-year 17 and 15-year 39 follow-up. Last, in patients with MS, lower cerebellar volumes were associated with poor motor and cognitive performance in a cross-sectional study. 40 Such results might reflect the extensive projections to the cerebellum from the limbs via the ventral and dorsal spinocerebellar tracts and from motor cortices through the middle cerebellar peduncles. 41 This study has some limitations. First, despite the large number of participants recruited, patients with CIS were less numerically represented than other phenotypes, probably reflecting their smaller prevalence in the general MS population. Moreover, the number of patients with progressive MS who completed the follow-up assessment was relatively small; therefore, we could not reliably calculate GM atrophy progression for PPMS and SPMS phenotypes separately. Third, patients with MS were older than HCs, possibly causing an enhancement of differences in GM volume between these 2 groups. However, statistical age adjustment was included in all statistical models. Fourth, the large majority of patients with RRMS received a DMT, while the opposite was true for patients with progressive MS. This numerical imbalance is probably the reason why we did not detect a significant influence of DMT on GM atrophy progression within phenotypes. Fourth, clinical worsening was assessed on a relatively short duration of follow-up (i.e., 1 year), which might not be sufficient to detect clinically relevant modifications in the majority of patients. Finally, neuropsychological assessment was not performed in our patients, preventing us from investigating correlations with cognitive impairment, depression, or fatigue symptoms.
SBM analysis revealed a differential involvement in various GM networks across disease stages, which progressed at 1 year in MS. Sensorimotor and cerebellar GM atrophy explained baseline disability and clinical worsening.