covarying patterns of white matter lesions and cortical ... · gm atrophy and wm lesions derived...

ARTICLE OPEN ACCESS

Covarying patterns of white matter lesions andcortical atrophy predict progression in early MSMuthuraman Muthuraman PhD Vinzenz Fleischer MD Julia Kroth MD Dumitru Ciolac MD

Angela Radetz MSc Nabin Koirala PhD Gabriel Gonzalez-Escamilla PhD Heinz Wiendl MD

Sven G Meuth PhD MD Frauke Zipp MDdagger and Sergiu Groppa MDdagger

Neurol Neuroimmunol Neuroinflamm 20207e681 doi101212NXI0000000000000681

Correspondence

Dr Groppa

segroppauni-mainzde

AbstractObjectiveWe applied longitudinal 3T MRI and advanced computational models in 2 independentcohorts of patients with earlyMS to investigate howwhite matter (WM) lesion distribution andcortical atrophy topographically interrelate and affect functional disability

MethodsClinical disability was measured using the Expanded Disability Status Scale Score at baselineand at 1-year follow-up in a cohort of 119 patients with early relapsing-remitting MS and ina replication cohort of 81 patients Covarying patterns of cortical atrophy and baseline lesiondistribution were extracted by parallel independent component analysis Predictive power ofcovarying patterns for disability progression was tested by receiver operating characteristicanalysis at the group level and support vector machine for individual patient outcome

ResultsIn the study cohort we identified 3 distinct distribution types of WM lesions (cerebellarbihemispheric and left lateralized) that were associated with characteristic cortical atrophydistributions The cerebellar and left-lateralized patterns were reproducibly detected in thesecond cohort Each of the patterns predicted to different extents short-term disability pro-gression whereas the cerebellar pattern was associated with the highest risk of clinical wors-ening predicting individual disability progression with an accuracy of 88 (study cohort) and89 (replication cohort) respectively

ConclusionThese findings highlight the role of distinct spatial distribution of cortical atrophy and WMlesions predicting disability The cerebellar involvement is shown as a key determinant of rapidclinical deterioration

These first authors contributed equally to this work

daggerThese senior authors contributed equally to this work

From the Department of Neurology (MM VF JK DC AR NK GG-E FZ SG) Focus Program Translational Neuroscience (FTN) Research Center for Immunotherapy (FZI)Rhine-Main Neuroscience Network (rmn2) University Medical Center of the Johannes Gutenberg University Mainz and Department of Neurology with Institute of TranslationalNeurology (HW SGM) University of Munster Germany

Go to NeurologyorgNN for full disclosures Funding information is provided at the end of the article

The Article Processing Charge was funded by the authors

This is an open access article distributed under the terms of the Creative Commons Attribution-NonCommercial-NoDerivatives License 40 (CC BY-NC-ND) which permits downloadingand sharing the work provided it is properly cited The work cannot be changed in any way or used commercially without permission from the journal

Copyright copy 2020 The Author(s) Published by Wolters Kluwer Health Inc on behalf of the American Academy of Neurology 1

MS is a chronic neuroinflammatory disease of the CNS thatleads to progressive disability in young adults Acute in-flammation and chronic neuronal cell loss contribute differ-ently to disease progression and disability12 Cortical graymatter (GM) atrophy starts at or before clinical disease onsetand the initial extent has robust associations with functionaldeterioration and cognitive impairment3 It reliably predictsand is a relevant marker of disease progression45 Along withcortical GM atrophy inflammatory demyelinating white mat-ter (WM) lesions are considered as primordial hallmarks of thedisease and their quantity at the beginning of the disease isassociated with future clinical outcome6 Furthermore lesionextent at disease onset is the most important conventional MRIfeature in the clinical setting that reliably predicts long-termdisability7

In the advanced stages of the disease GM and WM damagepatterns have been studied independently of each other andhave each been shown to be associated with motor dysfunc-tion or cognitive impairment8ndash10 Because both GM and WMdamage affect disease-related disability integrating analysis ofthese 2 compartments should provide more accurate pre-dictive models for disease progression

In our study we investigated distinct topological patterns ofcortical GM atrophy andWM lesions in patients with early MSand assessed the relevance of these patterns for disease pro-gression First we used parallel independent componentanalysis (ICA) to examine whether cortical atrophy and WMlesions derived from standardized 3T MRI follow specificpatterns in relation to each other This methodological ap-proach has recently been successfully implemented in the ex-ploration of GM andWM integrity in other neurodegenerativeand psychiatric disorders to depict patterns of specific tissueabnormalities1112 Then we investigated whether the identifiedpatterns were associated with disease progression in terms ofcortical atrophy and clinical disability after 1 year of follow-upTo confirm the reliability of our results we replicated theanalysis in an independent cohort of patients with MS

Previous studies investigatingGMandWMdamage commonlyused conventional lesion and atrophy measures capturing bothaspects of the disease inflammation and neurodegenerationbut did not take mutual pattern recognition analysis intoaccount13ndash15 Our hypothesis was that covarying patterns ofGM atrophy and WM lesions derived from structural MRI atdisease onset could accurately predict disability over time

MethodsStudy participants and designIn this longitudinal study a cohort of 119 patients with earlyrelapsing-remitting MS (RRMS) with a maximum disease du-ration of 5 years was included16 A diagnosis of RRMS wasestablished according to the 2010 revisedMcDonald diagnosticcriteria17 To verify the reliability of the obtained results fromthe main study cohort we replicated the analyses by includingan independent data set of a replication cohort of 81 patientswith RRMS from another medical center All patients un-derwent clinical evaluation including the Expanded DisabilityStatus Scale (EDSS) scoring and MRI examination and wereclinically and radiologically followed up after 12 monthsPatients were relapse- and steroid-free for at least 3 monthsbefore the MRI scans Patients were considered to have diseaseprogression if they had an increase of at least 1 point in theEDSS score at follow-up compared with baseline Patientshaving a decrease no increase or an increase of 05 in the EDSSscore were considered to have no progression18

Both MS cohorts were retrospectively analyzed from a pro-spective longitudinal observational study between 2011 and2016 comprising yearly follow-up assessment with a standard-ized protocol The minimal requirements for inclusion in our 2cohorts were a standardized MRI with concomitant clinicalassessment and a follow-up after 12 months (plusmn3 months)Dropout rates between both cohorts were comparable (studycohort 12 vs replication cohort 9)

Standard protocol approvals registrationsand patient consentsThe study was approved by the local medical ethics com-mittee (approval number 83754311 [8085]) all patientsprovided informed consent in accordance with the Declara-tion of Helsinki

MRI data acquisition

Study cohortImaging was performed on a 3TMRI scanner (MagnetomTimTrio Siemens Germany) with a 32-channel receive-only headcoil using a sagittal 3D T1-weighted magnetization-preparedrapid gradient-echo (MP-RAGE) sequence (echo timeinversion timerepetition time = 2529001900 ms flip an-gle = 9deg field of view = 256 times 256mm2 matrix size = 256 times 256slice thickness = 1 mm voxel size = 1 times 1 times 1 mm3) anda sagittal 3D T2-weighted fluid-attenuated inversion recovery

GlossaryAUC = area under the curve EDSS = Expanded Disability Status Scale FLAIR = fluid-attenuated inversion recovery FWHM =full width at half maximum GM = gray matter ICA = independent component analysis LST = lesion segmentation toolboxMNI = Montreal Neurological Institute MP-RAGE = magnetization-prepared rapid gradient-echo MSFC = MS FunctionalComposite ROC = receiver operating characteristic RRMS = relapsing-remitting MS SVM = support vector machineWM =white matter

2 Neurology Neuroimmunology amp Neuroinflammation | Volume 7 Number 3 | May 2020 NeurologyorgNN

(FLAIR) sequence (echo timeinversion timerepetition time= 38818005000 ms echo-train length = 848 flip angle =120deg field of view = 256 times 256 mm2 matrix size = 256 times 256slice thickness = 1 mm voxel size = 1 times 1 times 1 mm3)

Replication cohortPatients from the replication cohort were scanned with a 3TMRI (Magnetom Prismafit Siemens Germany) scanner withthe following acquisition parameters a sagittal 3D T1-weighted MP-RAGE (echo timeinversion timerepetitiontime = 229002130 ms flip angle = 8deg field of view = 256 times256 mm2 matrix size = 256 times 256 slice thickness = 1 mmvoxel size = 1 times 1 times 1mm3) and a sagittal 3D T2-weightedFLAIR (echo timeinversion timerepetition time = 38918005000 ms echo-train length = 250 flip angle = 120degfield of view = 256 times 256 mm2 matrix size = 256 times 256 slicethickness = 1 mm voxel size = 1 times 1 times 1 mm3)

In both cohorts the acquired T1-weighted images were thenprocessed by using an automated processing pipeline to extractcortical atrophy over 1 year and the T1- and T2-weightedFLAIR images for the lesion growth algorithm of WM lesionvolumes

MRI data processing

Quantification of WM lesion volumeThe baseline and follow-up volumes of WM lesions were es-timated by applying the cross-sectional pipeline of the lesionsegmentation toolbox19 as part of the Statistical ParametricMapping (SPM8) software Initially 3D FLAIR images werecoregistered to 3D T1-weighted images and bias correctedAfter partial volume estimation lesion segmentation was per-formed with 20 different initial threshold values for the lesiongrowth algorithm The optimal threshold (ĸ value dependenton image contrast) was determined for each patient and anaverage value for all patients calculated Afterward for auto-matic lesion volume estimation and filling of 3D T1-weightedimages a uniform ĸ value of 01 was applied in all patientsSubsequently the filled 3D T1-weighted images and the native3D T1-weighted images were segmented into GM WM andCSF and normalized to Montreal Neurological Institute(MNI) space Finally the quality of the segmentations wasvisually inspected to increase reliability

Cortical thickness reconstructionAfter filling of T1-weighted hypointense lesions the construc-tion of cortical surface and subcortical volume estimation foreach patient was performed based on 3D T1-weighted imagesusing FreeSurfer version 530 (surfernmrmghharvardedu) ina fully automated fashion followed by visual inspection forquality control at various processing steps Most of the patientsrsquoMRIs from both cohorts had sufficient quality and required noadditional intervention after initial automated processing Onlya few patients with excessive or insufficient skull stripping af-fecting the surface reconstruction were reran after manual res-toration of missing tissue or deleting nonbrain tissue Technicaldetails of the processing stream for surface-based reconstruction

were previously described and validated2021 Briefly this auto-mated processing pipeline is based on the creation of an un-biased within-subject template space and image using a robustinverse consistent registration20 The unbiased template servesfor initialization of skull stripping Talairach transformationatlas registration and parcellation for each subjectrsquos timepoints21 Surface maps of regional rates of cortical thicknessatrophy (inmmy) were computed as (thickness at time point 2minus thickness at time point 1)(time point 2 [years] minus time point1 [years]) and smoothed with a 10-mm full width at halfmaximum (FWHM) Gaussian kernel An appropriatesmoothing level increases the sensitivity of subsequent statisticalanalysis22 We used 10-mm FWHM because larger smoothinglowers the spatial resolution of cortical thickness measure-ments23 Subsequently within each of the detected lesion pat-terns differences in rates of cortical atrophy between the groups(classified according to the lesion volume of the pattern) wereevaluated Generated statistical parametric maps of significantgroup differences were corrected for multiple comparisons us-ing the false discovery rate (p lt 005) method

Identification of cortical GM atrophy and WMlesion patterns parallel ICAICA allows the identification of hidden noncorrelating com-ponents that underlie sets of measurements without any pre-vious knowledge of the data In this study we used the parallelICA algorithm implemented in the fusion ICA toolbox24 Thepurpose of parallel ICA is to discover independent componentsfrom 2 inputs here the rate of cortical atrophy and the volumesof baseline WM lesions derived from the MRI of the samepatient in addition to the relationship between them Beforeinputting into ICA cortical atrophy maps (registered in thehalfway space of the FreeSurfer longitudinal pipeline) weretransformed intoMNI space to ensure their fitting with theWMlesion maps The association between the influences of thecortical atrophy rate on the WM lesion volume was calculatedwith the algorithm by estimating the correlation between therate of cortical atrophy and theWM lesion volume Parallel ICAoptimization is based on the Infomax algorithm25 which max-imizes the mutual entropy to enhance the independence be-tween the components for the 2 inputs Finally to avoidoverfitting because of too many estimated parameters thelearning rate of the correlation term is adaptively adjusted26 Todetermine the correct number of independent componentsa modified version of the Akaike information criterion proposedby Li et al27 was applied The components from each modalitywere selected to be the correlations with the highest signifi-cance First we used the Akaike information criterion to esti-mate the number of components and then to reduce thecomponent number to reach a consistent level among the dif-ferent runs An overview of the analysis is illustrated in figure 1In a further analysis we used the Buckner functional cerebellarconnectivity atlas which parcellates the cerebellum into 7functional zones that are coupled to cerebral cortical regions(visual somatomotor dorsal attention ventral attention limbicfrontoparietal and default network)2829 This atlas depicts theorganization of cerebro-cerebellar circuits and delimits the

NeurologyorgNN Neurology Neuroimmunology amp Neuroinflammation | Volume 7 Number 3 | May 2020 3

intrinsic functional boundaries of the cerebellar-cortical net-works The previously obtained cerebellar pattern was projectedonto this atlas to identify specific functionally interacting cere-bellar networks The extent of the involvement of eachcerebellar-cortical network was quantified based on the absolutenumber of voxels (the threshold was set to gt10 voxels p lt 005)within each network and the relative size occupied by thesevoxels from the respective whole network

Individual-level support vectormachine analysisThe component loads of the lesion volumes for each individualpatient were determined from the identified significant cova-rying patterns These component loads for each lesion patternwere used as predictors of disease progression (as assessed bythe EDSS score at 1-year follow-up) for each individual patient

using a machine learning algorithm the support vector machine(SVM) The details of the applied SVM algorithm are presentedonly for the study relevant terms whereas the full information isavailable elsewhere30 In short the SVM algorithm is able toclassify 2 data sets based on an optimally separating thresholdbetween the data sets by maximizing the margin betweenclassesrsquo closest points The points which are located on theboundaries are called support vectors and the middle of themargin is the optimal separating threshold For an effectivelinear separation of data points a projection into a higher-dimensional space is required (here by means of polynomialfunction kernel) At the validation step the component loadswere assessed for the ability to automatically distinguish be-tween the patients with and without disability progression Theclassification was conducted for each lesion pattern separately

Figure 1 Data analysis flowchart

(A) WM lesion volumes derived from structuralMRI data sets (T1W T2W FLAIR) and (B) corticalatrophy rates (FreeSurfer processing) from thesame patients were used as inputs for the patternidentification analysis The weights were assignedbased on the correlation between these 2parameters Parallel ICA was performed and dis-tinct covarying lesion patterns of cortical atrophyand WM lesions were identified For each lesionpattern the component loads of each patient re-lated to cortical atrophy (brown line) and WM le-sion volume (green line) were correlated FLAIR =fluid-attenuated inversion recovery ICA = in-dependent component analysis T1W = T1‐weighted T2W = T2‐weighted WM = white matter


Statistical analysisSummary statistics are presented as mean plusmn SD median(range) or number (percentage) where applicable The pairedt test was used for parametric scale and the Wilcoxon signed-rank test for nonparametric scale variables All statistical anal-yses were performed with Statistical Package for the SocialSciences 200 software (IBM Armonk NY) and MATLABR2015b (MathWorks) To account for a possible influence oncortical atrophy rates age and sex of the patients were includedas covariates To investigate whether the detected lesion to-pology patterns allow the prediction of physical disability(measured with the EDSS) on the group level a receiver op-erating characteristic (ROC) analysis was conducted Thisstatistical method is preferentially used to quantify how accu-rately a diagnostic test performs when it is required to makea series of discriminations into 2 different states (eg stableEDSS and EDSS worsening) on the basis of a specific di-agnostic variable (eg lesion pattern) The area under the ROCcurve was used as the index of the prediction accuracy tocompare the ROC curves A p value of lt 005 was consideredstatistically significant For each significant covarying patternthat was used for the prediction of the EDSS score at follow-upthe accuracy was 10-fold cross-validated using the SVMalgorithm

Data availabilityThe raw data analyzed during this study will be shared inanonymized format by request of a qualified investigator tothe corresponding author for purposes of replicating proce-dures and results

ResultsPatientsDemographic clinical and MRI characteristics of the patientsfrom the study cohort are presented in table 1 The studycohort included 119 patients with RRMS (81 women) witha mean age of 346 plusmn 98 years and a mean disease duration of372 plusmn 250 months In total 104 patients (87) receiveddisease-modifying treatment and 15 patients (13) had nodisease-modifying treatment During the 1-year follow-up 12patients (10) presented clinical relapse 23 (19) had MRIactivity and 10 (9) presented both the remaining 74 patients(62) were clinically and radiologically stable Twenty-fivepatients (21) were considered to have disease progressionafter the 1-year follow-up period (maximum increase 10 in-crease in the EDSS score) At follow-up T2 lesion volume wassignificantly increased (t = 23 p lt 005) and global GM vol-ume was significantly decreased (t = 21 p lt 005) whencompared with baseline

The replication cohort comprised 81 patients (mean age 355 plusmn109 years 57 females) with a disease duration of 441 plusmn 105months (table 1) Within this cohort 50 patients (75) wereon disease-modifying therapy and 31 patients (25) were nottaking any disease-modifying medication In the course of

1-year period 9 patients (11) presented clinical relapse orMRI activity and 72 patients (89) were clinically and ra-diologically stable After 1 year 11 patients (14) had pro-gression in their EDSS scores (maximum increase 15 increasein the EDSS scores) At 1 year patients from the replicationcohort as those from the study cohort displayed reduced globalGM volume (t = 22 p lt 005)

Notably there were no significant differences in mean EDSSscores between baseline and follow-up neither in the studycohort (p = 051) nor in the replication cohort (p = 014) Bycomparison between the 2 cohorts (table e-1 linkslwwcomNXIA196) there were no significant differences in thepatientsrsquo clinical and MRI characteristics (all p values gt 005)except for disease duration being longer in the replication co-hort (372 vs 441 months p lt 0001)

Covarying patterns of WM lesions and regionalcortical atrophyThe WM lesion distributions and regional cortical atrophymaps determined from 119 MRI data sets of the study cohortwere included as 2 inputs into a parallel ICA The analysisshowed 3 significant covarying lesion patterns namely a cere-bellar pattern (r = 068 p lt 0001) a bihemispheric pattern (r =061 p lt 0001) implicating both hemispheres and a left-lateralized pattern (r = 058 p lt 0001) affecting pre-dominantly the left hemisphere (figure 2)

Similarly we juxtaposed the lesion and atrophy maps in thereplication cohort that were then fed into a parallel ICAHere thecorrelation analyses revealed 2 significant covarying lesion pat-terns cerebellar (r = 064 p lt 0001) with a similar lesion dis-tribution located bilaterally in the cerebellum and left-lateralized(r = 062 p lt 0001) involving mainly the hemisphere WM(figure e-1 linkslwwcomNXIA193)

Covarying regional atrophy representations were in additionanalyzed by subdividing the main cohort (n = 119) into 2subgroups dependent on the lesion load for each distinct pattern(below vs above themedian) These groups were comparable interms of global cortical atrophy rates (cerebellar [t = 135 p gt005] bihemispheric [t = 118 p gt 005] and left-lateralized [t =089 pgt005]) Regional atrophy distribution for each pattern ispresented in figure e-2 (linkslwwcomNXIA194) and tablee-2 (linkslwwcomNXIA196) None of the patterns was as-sociated with the global cortical atrophy (Pearson correlationanalysis p gt 005 for all patterns) in both cohorts

Covarying patterns predict diseaseprogression over timeWe performed ROC analyses (figure 3) to determine the pre-dictive ability of distinct patterns (GM and WM pathologycovariates) to discriminate patients with disability progression(based on patientsrsquo increase in the EDSS score by ge1 point atfollow-up) The area under the curve (AUC) was 082 (95CI073ndash091) specificity 79 and sensitivity 81 for the cerebellarpattern the AUCwas 066 (95CI 054ndash077) specificity 62


and sensitivity 59 for the bihemispheric pattern and the AUCwas 064 (95 CI 052ndash075) specificity 63 and sensitivity59 for the left-lateralized pattern (table e-3 linkslwwcomNXIA196) The cerebellar pattern therefore showed the bestpredictive performance of the 3 patterns in discriminating thepatients with and without disability progression over time Apairwise comparison of the ROC curves revealed a significantdifference between the AUC of the cerebellar pattern andbihemispheric (p lt 005) and left-lateralized (p lt 005) patternsrespectively

Analysis in the replication cohort supported these results Us-ing the ROC analysis we found that the AUC was 086 (95CI 076ndash094) specificity 83 and sensitivity 88 for thecerebellar pattern and the AUCwas 066 (95CI 055ndash077)specificity 61 and sensitivity 63 for the left-lateralized

pattern in this cohort (figure e-3 linkslwwcomNXIA195and table e-3 linkslwwcomNXIA196)

In both cohorts predictive accuracy derived from the in-tegration of both WM lesions and cortical atrophy patternsoutperforms the predictive accuracy of one of both patternsalone The comparative ROC characteristics are shown intable e-3 (linkslwwcomNXIA196)

The map of the cerebellar lesion distribution is shown in figure4A for the predictive power Here each particular voxel showshow this region contributes to the prediction analyses for dis-ease progression (patients with vs without disease pro-gression) This SVM-based individual patient predictionanalysis achieved as well the best cross-validated accuracy of88 for the cerebellar and of 75 for the bihemispheric and

Table 1 Demographic clinical and MRI-derived measures of the study and replication cohorts

Parameter Baseline Follow-up Ztb p Value

Study cohort

Demographic and clinicalcharacteristics

Age (y) 346 plusmn 98

Malefemale number () 38 (32)81 (68)

Disease duration (mo) 372 plusmn 250

EDSS scorea 15 (0ndash65) 15 (0ndash65) 06 051

Follow-up (mo) 120 plusmn 11

MRI characteristics

GM volume (mL) 6277 plusmn 589 6229 plusmn 550 21 004

WM volume (mL) 5689 plusmn 587 5686 plusmn 550 01 092

T2 lesion volume (mL) 73 plusmn 121 77 plusmn 123 23 002

Replication cohort







MRI characteristics




Abbreviations EDSS = Expanded Disability Status Scale GM = gray matter WM = white matterValues presented as mean plusmn SDSignificant p values (p lt 005) are shown in bolda Median (range)b Wilcoxon signed-rank (Z) or paired t test (t) between baseline and follow-up values


64 for the left-lateralized pattern in predicting disease pro-gression at 1-year follow-up (table e-4 linkslwwcomNXIA196) In the replication cohort the SVM-based individual

analysis achieved the best cross-validated accuracy of 89 forthe cerebellar and 66 for the left-lateralized patterns in pre-dicting the disease progression over 1 year (table e-4 linkslwwcomNXIA196)

While projecting the lesion maps of the cerebellar patternonto the Buckner connectivity atlas (figure 4 C and D) wedepict the involvement of 4 major cerebellar networks in bothcohorts (studyreplication cohort) namely frontoparietal(relative size range 50ndash27458ndash25) default mode(42ndash8260ndash80) somatomotor (31ndash5529ndash49)and ventral attention (28ndash6424ndash52) network

DiscussionIn this work we investigated covarying patterns of GM andWM tissue damage in patients with early RRMS and studiedthe impact of these patterns on MS disease progression Weidentified 3 covarying patterns of WM lesions and regionalcortical atrophy All 3 patterns predicted short-term diseaseprogression with clinical impairment robustly but with differ-ent accuracies The cerebellar pattern was related to a morewidespread mainly frontotemporal cortical atrophy and hadthe strongest prediction accuracy of the 3 patterns Because thecerebellum is functionally segregated and integrated in variousfunctional networks including interconnected subcortical andcortical areas a deafferentation-driven pathology secondary tolesions in cortico-ponto-cerebellar pathways could be postu-lated Our results suggest that cerebellar inflammation seems toamplify other pathologic mechanisms such as retrogradeneurodegeneration of cortical areas even far away from the

Figure 2 Three significant covarying patterns of regional cortical atrophy and the corresponding WM lesion volume

(A) Cerebellar lesion pattern (B) bihemispheric le-sion pattern and (C) left-lateralized lesion patternWM = white matter

Figure 3 Comparison of ROC curves showing the accuracyof covarying lesion patterns in predicting disabil-ity progression

The shadowed areas represent the 95 CIs In the legend the AUC and the95 CIs are presented The cerebellar pattern shows the best accuracy inpredicting disability progression over time in comparison to the bihemi-spheric and left-lateralized patterns (p lt 005) AUC = area under the curve


lesion site (as demonstrated by the covarying atrophy patternin our study) presumably via long-ranging connections thatwere recently shown to be more severely damaged than short-range connections31

Previous neuroimaging and histopathologic studies have shownthat both GM and WM compartments are affected early in thedisease course of MS3233 However it was still unclear whetherGMatrophy andWMdamage occur randomly are interrelatedor follow specific anatomical or sequential patterns and inparticular whether the topographical distribution is relevant forearly clinical disability Few new studies have addressed eitherthe patterns of WM damage8 or GM atrophy910 Steenwijket al9 depicted 10 different nonrandom patterns of corticalatrophy in patients with long-standing MS (disease durationgt20 years) 4 were associated with the functional disability(quantified by the EDSS score) at the time point of the in-vestigation The cortical atrophy patterns showed strongerassociations with clinical in particular cognitive dysfunctionthan global cortical atrophy The detected patterns that showed

most pronounced cortical atrophy were located in the bilateralposterior cingulate cortex and the bilateral temporal pole Inparticular themarked cortical atrophy of the temporal lobe is inline with our findings reporting that the cerebellar pattern ispredominantly related to atrophy of the superior temporal lobeand the cortical areas around superior temporal sulcus Incontrast to their findings we found that all our detected lesionpatterns comprisedmdashat least partlymdashcortical atrophy in theparietal lobe being responsible for sensory information amongvarious modalities including spatial sense and proprioceptionwhich is typically affected already in patients with early-stageMS as in our cohorts The study by Steenwijk et al provided thefirst evidence that spatial distribution of GM damage is relevantfor clinical disability inMS However it was limited by its cross-sectional nature clinically heterogeneous MS population(ie RRMS and primary and secondary progressive MS) andthe exclusion of subcortical and cerebellar GM

In our study we were able to identify 3 covarying patterns ofspatially distributed damage in a cohort consisting of only

Figure 4 Map of the cerebellar lesion distribution

Here each particular voxel shows how this regioncontributes to the prediction analysis for diseaseprogression in the study cohort (A) and in thereplication cohort (B) Color bar indicates thecomponent load of the cerebellar pattern (C)shows the Buckner 7-network cerebellar nucleiatlas and (D) represents the corresponding cor-tical connectivity map ICA = independent com-ponent analysis


patients with early RRMS who were evaluated in a longitudinalapproach Our findings indicate that these patterns are clinicallyrelevant and allow prediction of disease progression andemerging functional impairment in 2 early and mildly affectedRRMS cohorts We suggest that in early RRMS the patterns ofGM and WM pathology distribution reflect networkorganization3435 Our recent work has shown that such GMand WM network characteristics are essential for the mainte-nance of function through structural adaptation despite globallong-range disconnection3034 In particular at early stages theGM and WM display similar structural reorganization dy-namics with strengthening of local connectivity whereas atlater disease stages the divergence in network patterns is as-sociated with clinical deterioration3436

Our findings support the hypothesis that the distribution ofGM andWM damage is interrelated because we demonstratedcorrelative links between WM lesion patterns and localizationof cortical atrophy by parallel ICA We further showed that theglobal cortical atrophy was independent of the lesion patternphenotype and not related to clinical deterioration in the shorttime investigated here The cerebellar pattern shows most ex-tensive spreading of cortical atrophy and particularly withcortical thickness reduction in the frontal and temporal lobesThe bihemispheric pattern contains increased atrophy in theorbitofrontal cortex whereas the left-lateralized pattern cov-ered fusiform superior frontal and parietal cortical areas Thelack of an association between the lesion volume within eachpattern and the global rate of cortical atrophy highlights theimportance of the spatial distribution of cortical GM and WMdamage and shows that the simple notion that more tissuedamage necessarily leads to more clinical disability is too nar-row A recent neuropathologic study investigating the re-lationship between GM and WM damage postulated thatcortical inflammation and damage might drive the underlyingWM pathology through abnormal neuronal activity and de-creased input from the affected areas37 The authors providedevidence of direct interlesional connectivity between corticaland WM lesions in line with our findings that the interplay ofboth compartments influences disease progression

Although infratentorial pathology was reported to increase therisk of disability14 we here provide compelling evidence thatindividual-level SVM analysis of the cerebellar lesion pattern inconjunction with mainly frontotemporal cortical atrophy couldbe used to predict disability progression from the early diseasestages with high accuracy This could be used to build pre-diction models in an intersite setting and is of special clinicalrelevance for tracking disease courses or therapeutic responsesWe found the cerebellar pattern to be associated with higheratrophy rates in the superior temporal frontal and parietalcortex most of which is essential for motor and cognitiveperformance Indeed specific reorganization of cerebello-cortical networks is already evident in patients with clinicallyisolated syndrome and early RRMS30 and might represent anadaptive response to maintain motor function Breakdown ofthese networks on the other hand could lead to long-term

functional impairment or even be related to the transition intoprogressive form of the disease In addition the cerebellarcircuits undergo functional and structural reorganization withincreasing clinical disability3034 and are more prominent inprogressive MS3839

Some limitations apply to this work Although both cohortswere well matched on age sex and clinical disability thecohorts significantly differed in their disease duration How-ever the absolute difference of approximately 6 months be-tween both cohorts is clinically negligible as patients in bothcohorts are clearly in the early relapsing-remitting phase of thedisease Slight differences in some of the MRI acquisitionparameters between the 2 cohorts could have influenced thegeneration of the patterns although key acquisition parametersdetermining resolution were identical in both cohorts

Another limitation is the relatively short follow-up time of 1year as changes in the brain usually occur on a larger timescaleas changes on the EDSS score However cortical thinning asestimated in our study has shown to occur already early in thedisease and was repeatedly detectable in 1-year intervals40 Onthe other side the EDSS score is limited by its poor assessmentof upper limb function and cognitive decline which is capturedbetter in other outcome measures like the MS FunctionalComposite41 However despite its disadvantages it is still themost established score for evaluatingMS progression in clinicaltrials Finally only 21 within the main study cohort and 14within the replication cohort had proven disability progressionwhich is indeed in line with other 1-year progression data42 butresulted in a low absolute number of patients in our study andhence reduced the power of the statistical analysis

Our approach combining GM and WM imaging and corre-lating these with disease progression fills the gap in our un-derstanding between the functional consequences of structuraldamage in the cerebellum and their predictive value in patientswith MS Future work should evaluate and validate the long-term (over years and decades) implications of our findings andinvestigate the potential for a combined GM and WM bio-marker with broad clinical utility in patients with MS Identi-fication of MS imaging phenotypes with predominant signs ofcerebellar dysfunction would further affect early therapeuticstrategies potentially pointing to prevent a worse prognosis

Taken together we have shown that distinct damage distri-bution patterns in both GM and WM compartments ratherthan a broad random effect characterize upcoming disabilityprogression in the first year after onset of MS Lesion topologystrongly affects regional cortical integrity through inter-connected structural networks and accurately predicts theshort-term progression of individual patient disability Earlyrecognition and timely surveillance of lesion patterns at initialstages of the disease might be clinically applicable to identifypatients at higher risk of cortical atrophy and worse clinicaloutcome and who would therefore most benefit from prompttherapeutic intervention


AcknowledgmentThe authors thank Cheryl Ernest and Rosalind Gilchrist forproofreading the manuscript

Study fundingThis study has been supported by the German ResearchFoundation (DFG CRC-TR-128)

DisclosureMMuthuraman V Fleischer J Kroth D Ciolac A RadetzN Koirala G Gonzalez-Escamilla and S Groppa report nodisclosures relevant to the manuscript H Wiendl receiveshonoraria for acting as a member of scientific advisoryboards and as a consultant for Biogen Evgen MedDayPharmaceuticals Merck Serono Novartis Roche PharmaAG Sanofi Genzyme and speaker honoraria and travelsupport from Alexion Biogen Cognomed F Hoffmann-LaRoche Ltd Gemeinnutzige Hertie-Stiftung Merck SeronoNovartis Roche Pharma AG Sanofi Genzyme Teva andWebMD Global Prof Wiendl is acting as a paid consultantfor AbbVie Actelion Biogen IGES Novartis RocheSanofi-Genzyme and the Swiss Multiple Sclerosis SocietyHis research is funded by the BMBF DFG Else KronerFresenius Foundation Fresenius Foundation HertieFoundation NRW Ministry of Education and ResearchInterdisciplinary Center for Clinical Studies (IZKF)Muenster and RE Childrenrsquos Foundation Biogen GmbHGlaxoSmithKline GmbH Roche Pharma AG and SanofiGenzyme SG Meuth has received honoraria for lecturingtravel expenses for attending meetings and financial re-search support from Almirall Amicus Therapeutics GmbHDeutschland Bayer Health Care Biogen Celgene DiamedGenzyme MedDay Pharmaceuticals Merck SeronoNovartis Novo Nordisk ONO Pharma Roche Sanofi-Aventis Chugai Pharma QuintilesIMS and Teva F Zipphas recently received research grants andor consultationfunds from DFG BMBF PMSA MPG Genzyme MerckSerono Roche Novartis Sanofi-Aventis Celgene ONOand Octapharma Go to NeurologyorgNN for fulldisclosures

Publication historyReceived by Neurology Neuroimmunology amp NeuroinflammationJuly 11 2019 Accepted in final form January 7 2020

Appendix Authors

Name Location Role Contribution

MuthuramanMuthuramanPhD

University MedicalCenter of theJohannesGutenbergUniversity MainzGermany

Author Design andconceptualization ofthe study acquisitionand analysis of dataand draftingsignificantproportion of themanuscript

Appendix (continued)


VinzenzFleischer MD



Julia KrothMD



DumitruCiolac MD


Author Acquisition andanalysis of the dataand revision of themanuscript forintellectual content

AngelaRadetz MSc


Author Acquisition andanalysis of the data

Nabin KoiralaPhD



GabrielGonzalez-Escamilla PhD



Heinz WiendlMD

University ofMunster Germany

Author Analysis andinterpretation of thedata and revision ofthe manuscript forintellectual content

Sven GMeuth PhDMD



Frauke ZippMD



SergiuGroppa MD




References1 Siffrin V Vogt J Radbruch H Nitsch R Zipp F Multiple sclerosisndashcandidate

mechanisms underlying CNS atrophy Trends Neurosciences 201033202ndash2102 Zipp F Gold R Wiendl H Identification of inflammatory neuronal injury and pre-

vention of neuronal damage in multiple sclerosis hope for novel therapies JAMANeurol 2013701569ndash1574

3 Pitteri M Romualdi C Magliozzi R Monaco S Calabrese M Cognitive impairmentpredicts disability progression and cortical thinning inMS an 8-year study Mult Scler201723848ndash854

4 Fisher E Lee JC Nakamura K Rudick RA Gray matter atrophy in multiple sclerosisa longitudinal study Ann Neurol 200864255ndash265

5 Jacobsen C Hagemeier J Myhr KM et al Brain atrophy and disability progression inmultiple sclerosis patients a 10-year follow-up study J Neurol Neurosurg Psychiatry2014851109ndash1115

6 Galassi S Prosperini L Logoteta A et al A lesion topography-based approach topredict the outcomes of patients with multiple sclerosis treated with Interferon BetaMult Scler Relat Disord 2016899ndash106

7 Fisniku LK Chard DT Jackson JS et al Gray matter atrophy is related to long-termdisability in multiple sclerosis Ann Neurol 200864247ndash254

8 Meijer K Cercignani M Muhlert N et al Patterns of white matter damage are non-random and associated with cognitive function in secondary progressive multiplesclerosis Neuroimage Clin 201612123ndash131

9 Steenwijk MD Geurts JJ Daams M et al Cortical atrophy patterns in multiplesclerosis are non-random and clinically relevant Brain 2016139115ndash126

10 Bergsland N Horakova D Dwyer MG et al Gray matter atrophy patterns in multiplesclerosis a 10-year source-based morphometry study Neuroimage Clin 201817444ndash451

11 Gupta CN Chen J Liu J et al Genetic markers of white matter integrity in schizo-phrenia revealed by parallel ICA Front Hum Neurosci 20159100

12 Coppen EM van der Grond J Hafkemeijer A Rombouts SA Roos RA Early greymatter changes in structural covariance networks in Huntingtonrsquos disease Neuro-image Clin 201612806ndash814

13 Popescu V Agosta F Hulst HE et al Brain atrophy and lesion load predict long termdisability in multiple sclerosis J Neurol Neurosurg Psychiatry 2013841082ndash1091

14 Minneboo A Barkhof F Polman CH Uitdehaag BM Knol DL Castelijns JAInfratentorial lesions predict long-term disability in patients with initial findingssuggestive of multiple sclerosis Arch Neurol 200461217ndash221

15 Filippi M Preziosa P Copetti M et al Gray matter damage predicts the accumulationof disability 13 years later in MS Neurology 2013811759ndash1767

16 Chard D Miller D Grey matter pathology in clinically early multiple sclerosis evi-dence from magnetic resonance imaging J Neurol Sci 20092825ndash11

17 Polman CH Reingold SC Banwell B et al Diagnostic criteria for multiple sclerosis2010 revisions to the McDonald criteria Ann Neurol 201169292ndash302

18 Ellison GWMyers LW Leake BD et al Design strategies in multiple sclerosis clinicaltrials The Cyclosporine Multiple Sclerosis Study Group Ann Neurol 199436S108ndashS112

19 Schmidt P Gaser C Arsic M et al An automated tool for detection of FLAIR-hyperintense white-matter lesions in multiple sclerosis Neuroimage 2012593774ndash3783

20 Reuter M Rosas HD Fischl B Highly accurate inverse consistent registration a ro-bust approach Neuroimage 2010531181ndash1196

21 Reuter M Schmansky NJ Rosas HD Fischl B Within-subject template estimation forunbiased longitudinal image analysis Neuroimage 2012611402ndash1418

22 Pardoe HR Abbott DF Jackson GD AsDN Initiative Sample size estimates for well-powered cross-sectional cortical thickness studies Hum Brain Mapp 2013343000ndash3009

23 Coalson TS Van Essen DC Glasser MF The impact of traditional neuroimagingmethods on the spatial localization of cortical areas Proc Natl Acad Sci 2018115E6356ndashE6365

24 Calhoun VD Adali T Giuliani NR Pekar JJ Kiehl KA Pearlson GD Method formultimodal analysis of independent source differences in schizophrenia combininggray matter structural and auditory oddball functional data Hum Brain Mapp 20062747ndash62

25 Anemuller J Sejnowski TJ Makeig S Complex independent component analysis offrequency-domain electroencephalographic data Neural Netw 2003161311ndash1323

26 Liu J Pearlson G Windemuth A Ruano G Perrone-Bizzozero NI Calhoun VCombining fMRI and SNP data to investigate connections between brain functionand genetics using parallel ICA Hum Brain Mapp 200930241ndash255

27 Li YO Adalı T Calhoun VD Estimating the number of independent components forfunctional magnetic resonance imaging data Hum Brain Mapp 2007281251ndash1266

28 Buckner RL Krienen FM Castellanos A Diaz JC Yeo BT The organization of thehuman cerebellum estimated by intrinsic functional connectivity J Neurophysiol20111062322ndash2345

29 Yeo BTT Krienen FM Sepulcre J et al The organization of the human cerebralcortex estimated by intrinsic functional connectivity J Neurophysiol 20111061125ndash1165

30 Muthuraman M Fleischer V Kolber P Luessi F Zipp F Groppa S Structural brainnetwork characteristics can differentiate CIS from early RRMS FrontNeurosci 20161014

31 Meijer KA Steenwijk MD Douw L Schoonheim MM Geurts JJG Long-rangeconnections are more severely damaged and relevant for cognition in multiple scle-rosis Brain 2020143150ndash160

32 Kutzelnigg A Lucchinetti CF Stadelmann C et al Cortical demyelination and diffusewhite matter injury in multiple sclerosis Brain 20051282705ndash2712

33 Calabrese M Atzori M Bernardi V et al Cortical atrophy is relevant in multiplesclerosis at clinical onset J Neurol 20072541212ndash1220

34 Fleischer V Groger A Koirala N et al Increased structural white and grey matternetwork connectivity compensates for functional decline in early multiple sclerosisMult Scler 201723432ndash441

35 Fleischer V Radetz A Ciolac D et al Graph theoretical framework of brain networksin multiple sclerosis a review of concepts Neuroscience 201940335ndash53

36 Charalambous T Tur C Prados F et al Structural network disruption markersexplain disability in multiple sclerosis J Neurol Neurosurg Psychiatry 201990219ndash226

37 Tillema JM Weigand SD Mandrekar J et al In vivo detection of connectivity be-tween cortical and white matter lesions in early MS Mult Scler J 201723973ndash981

38 Anderson V Wheeler-Kingshott C Abdel-Aziz K et al A comprehensive assessmentof cerebellar damage in multiple sclerosis using diffusion tractography and volumetricanalysis Mult Scler J 2011171079ndash1087

39 Redondo J Kemp K Hares K Rice C Scolding N Wilkins A Purkinje cell pathologyand loss in multiple sclerosis cerebellum Brain Pathol 201525692ndash700

40 Koubiyr I Deloire M Coupe P et al Differential gray matter vulnerability in the 1Year following a clinically isolated syndrome Front Neurol 20189824

41 Rudick RA Cutter G Baier M et al Use of the multiple sclerosis functional compositeto predict disability in relapsing MS Neurology 2001561324ndash1330

42 Mesaros S Rocca MA Sormani MP Charil A Comi G Filippi M Clinical andconventional MRI predictors of disability and brain atrophy accumulation in RRMS Alarge scale short-term follow-up study J Neurol 20082551378ndash1383


DOI 101212NXI000000000000068120207 Neurol Neuroimmunol Neuroinflamm

Muthuraman Muthuraman Vinzenz Fleischer Julia Kroth et al early MS

Covarying patterns of white matter lesions and cortical atrophy predict progression in

This information is current as of February 5 2020

ServicesUpdated Information amp

httpnnneurologyorgcontent73e681fullhtmlincluding high resolution figures can be found at

References httpnnneurologyorgcontent73e681fullhtmlref-list-1

This article cites 42 articles 4 of which you can access for free at

Subspecialty Collections

httpnnneurologyorgcgicollectionmultiple_sclerosisMultiple sclerosis

httpnnneurologyorgcgicollectioncerebellumCerebellumfollowing collection(s) This article along with others on similar topics appears in the

Permissions amp Licensing

httpnnneurologyorgmiscaboutxhtmlpermissionsits entirety can be found online atInformation about reproducing this article in parts (figurestables) or in

Reprints

httpnnneurologyorgmiscaddirxhtmlreprintsusInformation about ordering reprints can be found online

Academy of Neurology All rights reserved Online ISSN 2332-7812Copyright copy 2020 The Author(s) Published by Wolters Kluwer Health Inc on behalf of the AmericanPublished since April 2014 it is an open-access online-only continuous publication journal Copyright

is an official journal of the American Academy of NeurologyNeurol Neuroimmunol Neuroinflamm

MS is a chronic neuroinflammatory disease of the CNS thatleads to progressive disability in young adults Acute in-flammation and chronic neuronal cell loss contribute differ-ently to disease progression and disability12 Cortical graymatter (GM) atrophy starts at or before clinical disease onsetand the initial extent has robust associations with functionaldeterioration and cognitive impairment3 It reliably predictsand is a relevant marker of disease progression45 Along withcortical GM atrophy inflammatory demyelinating white mat-ter (WM) lesions are considered as primordial hallmarks of thedisease and their quantity at the beginning of the disease isassociated with future clinical outcome6 Furthermore lesionextent at disease onset is the most important conventional MRIfeature in the clinical setting that reliably predicts long-termdisability7

In the advanced stages of the disease GM and WM damagepatterns have been studied independently of each other andhave each been shown to be associated with motor dysfunc-tion or cognitive impairment8ndash10 Because both GM and WMdamage affect disease-related disability integrating analysis ofthese 2 compartments should provide more accurate pre-dictive models for disease progression

In our study we investigated distinct topological patterns ofcortical GM atrophy andWM lesions in patients with early MSand assessed the relevance of these patterns for disease pro-gression First we used parallel independent componentanalysis (ICA) to examine whether cortical atrophy and WMlesions derived from standardized 3T MRI follow specificpatterns in relation to each other This methodological ap-proach has recently been successfully implemented in the ex-ploration of GM andWM integrity in other neurodegenerativeand psychiatric disorders to depict patterns of specific tissueabnormalities1112 Then we investigated whether the identifiedpatterns were associated with disease progression in terms ofcortical atrophy and clinical disability after 1 year of follow-upTo confirm the reliability of our results we replicated theanalysis in an independent cohort of patients with MS

Previous studies investigatingGMandWMdamage commonlyused conventional lesion and atrophy measures capturing bothaspects of the disease inflammation and neurodegenerationbut did not take mutual pattern recognition analysis intoaccount13ndash15 Our hypothesis was that covarying patterns ofGM atrophy and WM lesions derived from structural MRI atdisease onset could accurately predict disability over time

MethodsStudy participants and designIn this longitudinal study a cohort of 119 patients with earlyrelapsing-remitting MS (RRMS) with a maximum disease du-ration of 5 years was included16 A diagnosis of RRMS wasestablished according to the 2010 revisedMcDonald diagnosticcriteria17 To verify the reliability of the obtained results fromthe main study cohort we replicated the analyses by includingan independent data set of a replication cohort of 81 patientswith RRMS from another medical center All patients un-derwent clinical evaluation including the Expanded DisabilityStatus Scale (EDSS) scoring and MRI examination and wereclinically and radiologically followed up after 12 monthsPatients were relapse- and steroid-free for at least 3 monthsbefore the MRI scans Patients were considered to have diseaseprogression if they had an increase of at least 1 point in theEDSS score at follow-up compared with baseline Patientshaving a decrease no increase or an increase of 05 in the EDSSscore were considered to have no progression18

Both MS cohorts were retrospectively analyzed from a pro-spective longitudinal observational study between 2011 and2016 comprising yearly follow-up assessment with a standard-ized protocol The minimal requirements for inclusion in our 2cohorts were a standardized MRI with concomitant clinicalassessment and a follow-up after 12 months (plusmn3 months)Dropout rates between both cohorts were comparable (studycohort 12 vs replication cohort 9)

Standard protocol approvals registrationsand patient consentsThe study was approved by the local medical ethics com-mittee (approval number 83754311 [8085]) all patientsprovided informed consent in accordance with the Declara-tion of Helsinki

MRI data acquisition

Study cohortImaging was performed on a 3TMRI scanner (MagnetomTimTrio Siemens Germany) with a 32-channel receive-only headcoil using a sagittal 3D T1-weighted magnetization-preparedrapid gradient-echo (MP-RAGE) sequence (echo timeinversion timerepetition time = 2529001900 ms flip an-gle = 9deg field of view = 256 times 256mm2 matrix size = 256 times 256slice thickness = 1 mm voxel size = 1 times 1 times 1 mm3) anda sagittal 3D T2-weighted fluid-attenuated inversion recovery

GlossaryAUC = area under the curve EDSS = Expanded Disability Status Scale FLAIR = fluid-attenuated inversion recovery FWHM =full width at half maximum GM = gray matter ICA = independent component analysis LST = lesion segmentation toolboxMNI = Montreal Neurological Institute MP-RAGE = magnetization-prepared rapid gradient-echo MSFC = MS FunctionalComposite ROC = receiver operating characteristic RRMS = relapsing-remitting MS SVM = support vector machineWM =white matter





MRI data processing






























Study cohort







MRI characteristics




Replication cohort







MRI characteristics



































Appendix Authors







VinzenzFleischer MD



Julia KrothMD



DumitruCiolac MD



AngelaRadetz MSc



Nabin KoiralaPhD






Heinz WiendlMD



Sven GMeuth PhDMD



Frauke ZippMD



SergiuGroppa MD





























































Reprints







MRI data processing






























Study cohort







MRI characteristics




Replication cohort







MRI characteristics



































Appendix Authors







VinzenzFleischer MD



Julia KrothMD



DumitruCiolac MD



AngelaRadetz MSc



Nabin KoiralaPhD






Heinz WiendlMD



Sven GMeuth PhDMD



Frauke ZippMD



SergiuGroppa MD





























































Reprints




























Study cohort







MRI characteristics




Replication cohort







MRI characteristics



































Appendix Authors







VinzenzFleischer MD



Julia KrothMD



DumitruCiolac MD



AngelaRadetz MSc



Nabin KoiralaPhD






Heinz WiendlMD



Sven GMeuth PhDMD



Frauke ZippMD



SergiuGroppa MD





























































Reprints

































Appendix Authors







VinzenzFleischer MD



Julia KrothMD



DumitruCiolac MD



AngelaRadetz MSc



Nabin KoiralaPhD






Heinz WiendlMD



Sven GMeuth PhDMD



Frauke ZippMD



SergiuGroppa MD





























































Reprints





























































Reprints




covarying patterns of white matter lesions and cortical ... · gm atrophy and wm lesions derived...

Documents