Automatic Identification of Grey Matter Structures from MRI to Improve the Segmentation of White Matter Lesions Simon Warfield Joachim Dengler Joachim Zaers Charles R.G. Guttmann William M. Wells III Gil J. Ettinger John Hiller Ron Kikinis Abstract The segmentation of MRI scans of patients with white matter lesions (WML) is difficult because the MRI characteristics of white matter lesions are simi- lar to those of grey matter. Intensity based statistical classification techniques misclassify some WML as grey matter and some grey matter as WML. We developed a fast elastic matching algorithm that warps a reference data set containing informa- tion about the location of the grey matter into the ap- proximate shape of the patient’s brain. The region of white matter was segmented after segmenting the cortex and deep grey matter structures. The cortex was identified using a 3D region growing algorithm constrained by anatomical, intensity gradient and tis- sue class parameters. White matter and white mat- ter lesions were then segmented without interference from grey matter using a two class minimum distance classifier. Analysis of double echo spin echo MRI scans of sixteen patients with clinically determined multiple sclerosis (MS) was carried out. The segmentation of the cortex and deep grey matter structures provides anatomical context. This was found to improve the segmentation of MS lesions by allowing correct clas- sification of the white matter region despite the over- lapping tissue class distributions of grey matter and MS lesion. 1 Introduction The segmentation of MRI scans of patients with white matter lesions is difficult because the MRI characteristics of white matter lesions (WML) are similar to those of grey matter. In- tensity based statistical classification techniques misclassify some WML as grey matter and some grey matter as WML. We developed a technique to identify a mask of the white matter region of the brain. White matter and white matter lesions were then segmented without interference from grey matter. The use of MRI for monitoring of treatment trials in MS has been analyzed. 9 The characterization of MS requires long term serial studies (typically 1-2 years) because of the relatively slow progression of the disease. MRI scans with thin contiguous slices are desirable, with scanning repeated at intervals of 1-3 months. Quantitative assessment of the ap- pearance of high intensity regions in MRI scans is important to evaluate the disease activity and progression. 12 Manual analysis of this volume of data is expensive and tedious. The intra- and inter-rater reliability of semi-automatic methods of analysis range from 5-20%. 9 The requirements to reduce human interaction (in order to improve reproducibility and to derive a measure of lesion burden that is independent of the operator) and to improve the accuracy of lesion load mea- surements were identified as important goals for any new segmentation methods in a recent review. 4 However, auto- matic segmentation is difficult because of the similarity of the pixel intensity of MS lesions and grey matter. Attempts have been made to deal with the overlapping intensity range of normal grey matter and lesion tissue. One approach used a model involving a spatially varying prior probability density for brain tissue class. 7 The search for MS lesions was confined to those regions with at least a 50% prior probability of being white matter. In this way, the in- correct classification of grey matter as MS lesion was greatly reduced. This model compensates for the tissue class inten- sity overlap by using a probabilistic model of the location of MS lesions. Our method segments the white matter region from each scan, rather than using a probabilistic model for [email protected]. Journal Image Guided Surgery, Vol. 1, Num. 6, pp.326–338, 1995, and at http://journals.wiley.com/cas/v1n6/95042-intro.html Vision Systems, Neckargemuend, Germany, [email protected]DKFZ (German Cancer Research Center) D-69120 Heidelberg Germany, [email protected]Harvard Medical School and Brigham and Women’s Hospital, Department of Radiology, 75 Francis St., Boston, MA 02115, [email protected]Harvard Medical School and Brigham and Women’s Hospital, Department of Radiology, 75 Francis St., Boston, MA 02115, [email protected]AI Lab NE43-771, Massachusetts Institute of Technology, 545 Technology Square, Cambridge MA 02139, [email protected]UCLA School of Medicine, Department of Radiological Sciences, [email protected]Harvard Medical School and Brigham and Women’s Hospital, Department of Radiology, 75 Francis St., Boston, MA 02115, [email protected]
white matterlesions(WML) is difficult becausetheMRI characteristicsof white matterlesionsaresimi-lar to thoseof grey matter. IntensitybasedstatisticalclassificationtechniquesmisclassifysomeWML asgrey matterandsomegrey matterasWML.
We developeda fast elasticmatchingalgorithmthat warpsa referencedataset containinginforma-tion aboutthelocationof thegrey matterinto theap-proximateshapeof the patient’s brain. The regionof white matterwassegmentedaftersegmentingthecortex anddeepgrey matterstructures.The cortexwasidentifiedusinga 3D region growing algorithmconstrainedby anatomical,intensitygradientandtis-sueclassparameters.White matterandwhite mat-ter lesionswerethensegmentedwithout interferencefrom grey matterusingatwo classminimumdistanceclassifier.
Analysisof doubleechospinechoMRI scansofsixteenpatientswith clinically determinedmultiplesclerosis(MS) wascarriedout. Thesegmentationofthe cortex anddeepgrey matterstructuresprovidesanatomicalcontext. This wasfound to improve thesegmentationof MS lesionsby allowing correctclas-sificationof thewhitematterregiondespitetheover-lappingtissueclassdistributionsof grey matterandMS lesion.
We developeda techniqueto identify a maskof the whitematterregion of the brain. White matterandwhite matterlesionswerethensegmentedwithout interferencefrom greymatter.
Theuseof MRI for monitoringof treatmenttrials in MShasbeenanalyzed.9 The characterizationof MS requireslong termserialstudies(typically 1-2 years)becauseof therelatively slow progressionof the disease.MRI scanswiththin contiguousslicesaredesirable,with scanningrepeatedatintervalsof 1-3months.Quantitativeassessmentof theap-pearanceof high intensityregionsin MRI scansis importantto evaluatethe diseaseactivity andprogression.12 Manualanalysisof thisvolumeof datais expensiveandtedious.Theintra- and inter-rater reliability of semi-automaticmethodsof analysisrangefrom 5-20%.9 Therequirementsto reducehumaninteraction(in orderto improvereproducibilityandtoderive a measureof lesionburdenthat is independentof theoperator)andto improve the accuracy of lesion load mea-surementswere identified as importantgoalsfor any newsegmentationmethodsin a recentreview.4 However, auto-matic segmentationis difficult becauseof the similarity ofthepixel intensityof MS lesionsandgrey matter.
Attemptshave beenmadeto dealwith the overlappingintensityrangeof normalgrey matterandlesiontissue.Oneapproachuseda model involving a spatially varying priorprobability densityfor brain tissueclass.7 The searchforMS lesionswasconfinedto thoseregionswith at leasta50%prior probabilityof beingwhite matter. In this way, the in-correctclassificationof grey matterasMS lesionwasgreatlyreduced.This modelcompensatesfor thetissueclassinten-sity overlapby usingaprobabilisticmodelof thelocationofMS lesions. Our methodsegmentsthe white matterregionfrom eachscan,ratherthanusinga probabilisticmodelfor
andathttp://journals.wiley.com/cas/v1n6/95042-intro.htmlVisionSystems,Neckargemuend,Germany, [email protected] (GermanCancerResearchCenter)D-69120Heidelberg Germany, [email protected]�Harvard Medical School and Brigham and Women’s Hospital, Department of Radiology, 75 Francis St., Boston, MA 02115,
Elastic matching algorithms are not able to segmentstructuresin the patient that are not presentin the atlas,suchaswhite matterlesions. Unlike somebrain tumours,MS lesionsdo not causesignificantdistortionof a patient’sanatomy. Elasticmatchingmaybeunableto compensateforthedistortioncausedby atumour, becausethedifferencebe-tweentheatlasandpatientbrainstructureshapeis not welldescribedby an elasticdeformation. The normalanatomi-cal variability of thecortex is alsonot well describedby anelasticdeformation.
We have developeda segmentationmethodthatusesthepositive featuresof both statisticalclassificationand elas-tic matchingmethodsto overcomethe limitations presentwhenthetechniquesareusedalone.Ourmethodis ageneralmethodwhich usesanatomicalinformationto disambiguatethesegmentationwhenclassdistributionsoverlapin featurespace.This approachis well suitedto the segmentationofMRI scansof MS patientsbecauseof thesubstantialdegreeof overlapof the MS lesionclasswith other tissueclasses(particularlygrey matter).It is a new automaticmethodthatproducesasegmentationthatis superiorto thatpossiblewitheitherelasticmatchingor statisticalclassificationalone.
The principal resultsof this studyarean algorithmforthe automaticsegmentationof the cortex, a methodfor theautomaticidentificationof the region of white matterandamethodfor thesegmentationof WML. Thismethodhasbeenappliedto MRI scansof sixteenpatientswith clinically de-terminedMS andto scanswith simulatedlesions.Togethertheseindicatethat the methodaccountsfor the ambiguitydueto theoverlappingintensitydistributionsof grey matterandWML, andconsequentlyimprovesthesegmentationofthesetissueclasses.
2 Description of Method
Moti vation
A majordifficulty for accuratesegmentationof WML is theoverlappingintensitydistributionsof WML andgrey mat-ter. Someregions of lesion cannotbe distinguishedfromgrey matterand vice versaeven when powerful nonpara-metric multispectralstatisticalclassificationtechniquesareused.Somelesionshave intensitycharacteristicsentirely inthe rangeof that of grey matter, with no voxel having anunambiguouslesion intensity. They can be recognizedaslesionby an experiencedobserver asthey appearin the re-gion of the white matterwhereno grey matteris expected.The segmentationmustbe donein 3D since2D slicescanshow small blobsof cortex which appearto be white mat-ter lesions.In a slicetheseappearisolated,but examinationof neighbouringslicesshows that they areconnectedto thecorticalmantle.
Thesameproblemmanifestsitself in anotherform. Thetypical slicethicknessusedin MS studieswith doubleechospin echoMRI gives rise to significantpartial volume ar-tifacts. The relatively low spatialresolutionof thesescansgivesrise to boundariesbetweendifferenttissuetypesthatare difficult to distinguish. This occursparticularlyat theedgeof thebrain,wherecerebrospinalfluid (CSF)andgreymatterareaveragedtogether. This sometimescausesvoxelsto havea pixel intensityrangetypicalof MS lesion.Any in-tensitybasedstatisticalclassifierwill misclassifysuchvox-elsasMS lesion.
If it is assumedthat MS lesionsalways occur withinwhite matterthenwe candifferentiatebetweenMS lesionsin whitematterandothervoxelswith pixel intensitiessimilarto MS lesionsusingournew segmentationalgorithm.
SegmentationAlgorithm Overview
Thesegmentationalgorithmis representedschematicallyinFigure1. Theprimary input to thesegmentationprocessisa doubleechospin echoMRI scanof a patient. The MRIscanis smoothedto improve SNRandthe intracranialcav-ity is segmented.Classificationandintensityinhomogeneitycorrectionis carriedout with theExpectation-Maximizationalgorithm. Our atlasis first alignedto thepatientscanwitha six parameter(threetranslationandthreerotation) lineartransformation,andthenlocalshapedifferencesareresolvedwith our elasticmatchingalgorithm. Thedeepgrey matterstructuresaresegmenteddirectlyby elasticmatching.How-ever, thenormalanatomicalvariability of thegrey matteroftheneocortex is notwell describedbyanelasticdeformation.Consequently, a new 3D constrainedregion growing algo-rithm wasdevelopedin orderto segmentthe cortex. Afterthegrey matterstructureshave beensegmentedusingthesealgorithms,thewhitematterregion is segmentedandreclas-sified into white matterandWML classes.This processre-
Edgepreservingnoisesmoothingis carriedout on thescanby iteratively solvingthenonlineardiffusionequation������������������ �! where
�is the vector-valuedMRI scan,
�is time (iteration
number), �"� ##%$'&)( *,+-(.0/21 is the spatially varying conduc-
tancefunctionand 3 is thenoisestrengthparameter.5 Sincethe in-planeresolutionof our scans(0.9375mm)is muchsmallerthanthe thicknessof theslices(3mm), theconduc-tanceacrosstheslicesof thescansis negligible. It hasbeenshownthattheuseof nonlineardiffusionfiltering reducesthevariability of operator-guidedsegmentationof MS lesionsacquiredwith 1.5T MRI exams.10
Intracranial Cavity Segmentation
The intracranialcavity (ICC) is segmentedusing a semi-automaticmethod.8 A parzenwindow classifieris usedtosegmentthevolumeinto brainandnon-brainclasses.SomematerialoutsidetheICC is still classifiedasbrain,so3D ero-sionfollowedby supervisedconnectivity analysisand3D di-lation is performed.Theresultingbinarydatasetrepresentsthe region of the intracranialcavity. This is the only pro-cessingstepthatrequiresmanualinteractionfor eachpatientvolume. Theinterraterreliability hasbeenassessedandthestandarddeviation of theareaof the ICC segmentationwasfoundto be0.5%.8 TheICC maskis usedfor thecalculationof the linearregistrationandto excludematerialexternaltothebrainfrom thelaterprocessingsteps.
Intensity-basedstatistical classificationand intensity in-homogeneitycorrectionare calculatedsimultaneouslyus-ing theExpectation-Maximization(EM) segmentationalgo-rithm.17 TheEM segmentergeneratestwo datasetsfor eachpatientdataset. Theseare the statisticalclassificationofeachvoxel and the valueof the DE SE MRI after correc-tion for the intensityinhomogeneity. Automaticcorrectionof intensityinhomogeneityin thisway makesthesegmenta-tion robustagainstminor changesin MR operatingcharac-teristicsaswell ascorrectingfor theinherentintensityinho-mogeneity(shadingartifact). Consequently, it is possibletocarry out supervisedclassificationon onepatientscanandto usethe samestatisticalmodelfor the distribution of tis-sueclassesto segmentotherscansof the sameacquisitiontype.Thereis noneedto retraintheclassifier. Thisalgorithmis run fully automatically, without any userintervention,oneachnoisesmoothedpatientscan.
Thealgorithmproceedsby iteratively estimatingthetis-sueclassand intensity inhomogeneityat eachvoxel, until
both estimateshave converged. The intensityinhomogene-ity is modelledasanadditivebiasfield in a log-transformeddatasetandis estimatedby solving457698 ��:<;>=�? where
45is a vectorrepresentingthe biasfield,
8is a low
passfilter, : is a vector of the observed tissueintensities(imagepixel values),?@�A�CBD�E� # GF BD�E�GH GF�I2I�IJ is a vectorofthe meanintensityof eachtissueclass,�2K representstissueclassk and = is a matrix of a posterioritissueclassproba-bilities (weightsusedto predictthesignalintensity).
The distribution of observed pixel valuesfor a singlechanneldatasetis modelledbyL �E:,MON PQM F 5 M �SRUTV�E:,MV;WBD��PQM ; 5 M whereR T ��X is azeromeanGaussiandistributionwith vari-anceY H , andthetissueprobability(weight)matrix = MJZ canbeestimatedby
= M[Z 6 \]�E�^Z L ��:�M�N �^Z F 45 M _a`)b \]��P M L ��: M N P M F 45 M where \]�E� Z is the a priori probability of tissueclass � Z .Voxel c is classifiedby selectingtheclassfor which = M[Z islargest.
Linear Registration
A lineartransformation(with threetranslationandthreero-tationparameters)is computedin orderto align the ICC oftheatlaswith thatof thepatient.Thegoalof the linearreg-istration is to computea global alignmentof the atlasandpatientdatasetsso that the elasticmatchingalgorithmthatfollowscansuccessfullydeformtheatlasto accountfor localshapedifferences.Thelineartransformationis computedbyminimizing thedistancebetweenthesurfacesof theICCs.
The voxels at the surfaceof the ICC for the atlasandpatientaresegmented.For eachslice,the largestconnectedcomponentof thebackgroundis foundandthevoxelsof theICC adjacentto thisareidentifiedasthesurfacevoxels.Thelinear transformationd is found by usingPowell’s methodto minimize e M0fhgji kml Hnpo�q F fhgjiZ N d'r M�; L Z�N swhere r M aretheatlassurfacevoxels, L Z arethepatientsur-facevoxels and
l npo�qis a maximumdistancethresholdto
limit thecontributionany singlevoxel canmake. TheRMSerror in thematchobtainedwith this methodis of theorderof thevoxel size.3
3
Elastic Matching
A 3D volumetricreferencedataset(atlas)is matchedto theanatomyof thepatientin orderto providea segmentationofthe normalanatomy. Our anatomicalatlaswasconstructedby a combinationof statisticalclassificationandmanualla-belling of every voxel in a 3D SPGRMRI scanof a nor-mal volunteer. It consistsof both a grey scaledatasetanda labelleddataset.16 Thepurposeof theatlasis to describenormalbrain anatomyfor a specificpersonin sucha waythat the atlaslabelscan be transferredto anotherperson’sanatomy, without the needto repeatthe arduousandtime-consuminghandlabellingthatwasrequiredto constructtheatlas.We have developeda fastandeffective elasticmatch-ing algorithmby which the atlaslabelscan be transferredautomaticallyto anotherdataset.2
Thenonlineartransformationrelatingtheatlasanatomyto the patient anatomyis modelledas a 3D deformationfield. Estimationof the local displacementvector ?t���u F-v,F-wx %y �z?]�CX F-{�F-|! can be formulatedas a standardregularizationproblem.14 Thedeformationis determinedbyminimizing} �~? ���h�h��\]�~? ����,� ��? l X l { l |This functional describesa balancebetweenthe deforma-tion energy
� �~? andlocal similarity of the images\]��? asa functionof thedeformation.Thelocal similarity \]��? is determinedby binary correlationof the atlasandpatientsegmenteddatasetsoverasmallneighbourhood.Thedefor-mationenergy term
� �~? is basedon a physicalmodelof a3D elasticmembrane,andis independentof thedata.This is� ��? �0u Hq � u H� � u H� ��v Hq �>v H� ��v H� �>w Hq ��w H� ��w H�
Applying a finite elementdiscretizationleadsto linearEuler-Lagrangeequationscorrespondingto the functional.This systemof equationscanbe efficiently solved usinganestedmultigrid algorithm with conjugategradientrelax-ation.15
The deformationfield is calculatedin two stages.Ini-tially it is assumedto beflat. ? is estimatedon thebasisofthebrainparenchyma(grey matterandwhitematterandforthe purposeof matching,WML is treatedaswhite matter).This is doneby solving for ? usingbinary datasetscorre-spondingto thebrainparenchymafrom theatlaslabelsandfrom the classifiedpatientdataset. The deformationfieldobtainedis thenusedasthe initial deformationandthe es-timateis refinedby repeatingthematchingprocedureusingbinarydatasetscorrespondingto only whitematter.
Brainstructuresthathavea regularandconsistentshapeacrossmany people,suchasthedeepgrey matterstructuresof the diencephalonandtelencephalon,andthe cerebellumaredirectly segmented. Somedeepgrey matterstructures(particularlytheglobuspallidus)canhaveanintensitychar-acteristicsimilar to that of white matter, and so are mis-classifiedby intensitybasedstatisticalclassification.Elastic
However, thevariationof theshapeof thecortex betweendifferentpeopleis not well describedby anelasticdeforma-tion. The atlasrepresentsthe anatomyof onenormalvol-unteer. Thenormalanatomicalvariability of thecortex canmeanthata patienthasa gyruswheretheatlashasnone,orvice versa.Whenelasticmatchingis carriedout, theregionof thecortex is recovered,but preciselocalizationof thebor-dersof thecortex is notachieved.
Segmentationof the Cortex
Segmentationof thecortex hasnot beenachievedusingei-ther statisticalclassificationor elasticmatchingtechniquesalone. Statisticalclassificationis able to identify the greymatterthatmakesup thecortex, but not to differentiatebe-tweenthe cortex and othergrey matterstructures.Elasticmatchingprovidesan estimateof the region of the cortex,but nota preciselocalizationof theboundaryof thecortex.
We havedevelopeda new algorithmfor segmentationofthecortex with a constrained3D region growing algorithm.The algorithmusestheanatomicalinformationprovided intheatlas,tissueclassinformationfrom statisticalclassifica-tion, the dataintensitygradientanda simplemodelof thestructureof thecortex.
Model of the cortex
We assumethat thecortex hastheform of a thick “blanket”crumpledover itself. Oneedgeof thecortex, theouteredge,is adjacentto CSF. Thereexistsa pathjoining every pair ofvoxels in the cortex that passesonly throughvoxels of thecortex. Therefore,it is possiblefor a region growing algo-rithm to segmentthecortex. It is necessaryto considerthestructureof thecortex in 3D, sincein 2D slicespartsof thecortex mayappearto bedisconnectedfrom otherparts.
Selectionof cortex seedvoxels
Theregiongrowing algorithmis startedfrom seedlocationsautomaticallyidentifiedat theouteredgeof thecortex. Theregion outsidethe ICC is maskedout of theclassifiedMRIdatasetandCSFis removedby relabellingit asbackground.At theouterboundaryof thecortex thereis oftena layerofmisclassifiedvoxels,causedby partialvolumeaveragingofthe brain andCSF. This is removedby performingerosionon eachslice of the classifieddataset with a 3x3 circularstructuringelement.Voxelson theoutersurfaceof thebrainareidentifiedwith thesamealgorithmthatis usedto identifythesurfaceof theICC for linearregistration.Theatlaslabelsareusedto identify andeliminatevoxelsthatareoutsidetheregionof thecortex. Thelocationof eachvoxel identifiedisthenusedasa seedlocationfor theregiongrowing.
4
Anatomical constraint
Theelasticmatchof theatlasto thecortex providesa goodestimateof the region of the cortex but doesn’t accuratelylocalize the bordersof the cortex. A region that includesthebordersof thecortex is generatedby 3D dilation of thematchedcortex. This new region formsa maskthat is usedto restricttherangeoverwhichregiongrowing is carriedout.Themaskrepresentsanestimateof themaximumerror thatis possiblein theelasticmatchof thecortex.
The deepgrey matterstructuresandthe cerebellum,asdeterminedby theelasticmatch,aremaskedoutsothatthesestructurescan’t beincorrectlyincludedin thecortex region.
Tissueintensity constraint
The cortex is a structureof grey matter. Only voxels withthepixel intensityrangeof grey matter, asdeterminedwiththeEM segmentationalgorithm,areconsidered.Partial vol-umeaveragingartifactsarecorrectedby relabellingafterthecortex is segmented.
Tissuegradient constraint
White matter lesionscan appearas a bright centreregionwith adimmer“halo” surroundingit andalsoasabright rimsurroundingadarkercentre.6 This impliesthatthereis asig-nificantrateof changeof intensityaswe move out from thecentreof a lesion. A high rateof changeof pixel intensityis indicative of an edgein the image. An edgeis foundbylocatinglocal maximaof thegradientafter smoothingwithnonlineardiffusion.13 Theregiongrowing processis not al-lowed to crossan edge. This constraintis implementedbyconstructinga binary volumein which voxels areset if anedgeis indicated,andarenot setin theabsenceof anedge.This constraintis usedsothat theregion will not beabletogrow fromthecortex regiontovoxelsaroundalesionif thereis anintensitygradientpresent.
Cortex SegmentationRegionGrowing Algorithm
The location of eachseedvoxel is placedin a queueandmarkedin a binaryvolumerepresentingthecortex. Regiongrowing proceedsby applyingthefollowing until thequeueis empty, creatingabinaryvolumerepresentingthesegmen-tationof thecortex.
Take locationv from headof queue,For eachl in theneighbourhoodof v,
if (l is notmarkedascortex) andif (l is in thecortex mask)andif (l hasgrey matterclass)andif (l is notata peakin theintensitygradient)then
Partial volumeaveragingcausessomegrey matterto beincorrectly classifiedas white matterlesion. This error isparticularlysevere in the slicesat the top of the brain andaroundthecerebellum.Theregionsattheboundarybetweenthecortex andCSFwhichareclassifiedaslesiondueto par-tial volumeaveragingarecorrectedby relabellingthesere-gionsasgrey matter.
The white matterregion is thensegmented.The whitematterregion is theregion insidetheintracranialcavity, lessthosevoxels identifiedasCSF, lessthosevoxelssegmenteddirectly by elasticmatching,lessthosevoxels identifiedascortex.
Segmentationof White Matter Lesions
Somevoxelsin thewhitematterregionareerroneouslyclas-sified as grey matterby the EM segmenterbecauseof theoverlappingintensityrangeof thegrey matterandthewhitematterlesionclasses.Voxels in this region shouldbe clas-sified only aswhite matteror aswhite matterlesion. Eachvoxel in theregionis reclassifiedaseitherwhitematteror aswhitematterlesionusinga two classtwo channelminimumdistanceclassifier.
3 Results
Double echo spin echo MRI scans of sixteen patientswith clinically determinedmultiple sclerosiswereacquiredon a GE Signa 1.5T clinical scannerwith TR/TE# /TEH3000/30/80ms, FOV 24cmandcontiguous3mm thick ax-ial slicescoveringtheentirebrain.
Thesegmentationprocesswasappliedto eachMRI scan.The segmenteddatasetswere then analysedto assesstheconsistency andaccuracy of thesegmentationof thecortexandMS lesions.
The EM segmentationalgorithm and rigid registrationeachrequiredaboutonehour to executeon eachscan,run-ning on a SunSPARCstation20/612.The3D elasticdefor-mationwascalculatedonanIBM RS/6000workstation.Thecalculationrequiredabout7 minutesto projectthe3D volu-metricatlasdatasetontoaMS patientdataset.Theaccuracyof thematchwasimprovedby iteratingtheelasticmatchpro-cedure5 times.Thecortex andlesionsegmentationrequiredabout30 minutesperscanon a SunSPARCstation20/612.Thesegmentationof atypicaldatasetrequiredabout3 hoursoverall.
To assessthe effectivenessof our cortex segmentationalgorithm,we comparedthesegmentationwith thesegmen-
Table1: Comparisonof raterperformancewith “standard”segmentationof the cortex in oneslice. Agreementis mea-suredasthepercentageof voxelscorrectlysegmented— �2��� �2�a�����!�Q����
whereS is a segmentationandC is the“standard”segmentationformedby selectingthosevoxelsthatat least5 of the6 segmentationsagreewerecortex.
tation of five different raters. The cortex was segmentedonceby five differentratersandby our automaticanalysismethod. The raters“painted” the region of the cortex onthe sameslice of onepatientscan. Sincethe true segmen-tation of the cortex is unknown, a “standard”segmentationof the cortex for comparisonwas constructedby selectingthosevoxelsthatat leastfiveof thesix segmentationsagreedwere in the cortex. The agreementof a segmentationwiththe standardcortex segmentationwasmeasuredasthe per-centageof voxelsof thestandardcortex correctlysegmented.Table1 shows theagreementof theratersegmentationsandtheautomaticsegmentationwith thatof the“standard”cor-tex segmentation.
Figure2(a) illustratesthevariationin segmentationthatdifferent ratersproduce. Figure 2(e) shows the region se-lectedasthe standardcortex segmentation.The segmenta-tionsof differentratersareconsistentin the labellingof thecentreregion of the cortical grey matter, but vary in deter-mining theinnerandouterboundaryof thecortex. This re-sult is consistentwith theobservationthatstructureswith acomplex boundaryshapearemoredifficult for peopleto seg-mentthanstructureswith a simpleboundary.8 Peoplehavea tendency to over- or under-estimatetheboundary, ascom-paredto anautomaticsegmentation.Consequently, manualsegmentationexhibitsaconsistentvariability in thesegmen-tationof voxelsat theboundaryof corticalgrey matter.
Thesegmentationof simulateddatasetswascarriedoutin order to evaluatethe segmentationwhen the true seg-mentationis known. A lesionwith a bright centralregionsmoothlydimmingtowardstheborderwasselectedfromonesliceof a patient’sMRI scan.Statisticalclassificationof thelesion incorrectlysegmentsit into a lesioncomponentanda grey mattercomponent.This lesionwasinsertedinto thetestslice andthe entirescanwassegmented.The segmen-tation of the lesionwasfound to vary with the lesionloca-tion. Whentheentirelesionwasplacedinto thewhitematterit wassegmentedaccurately. Whenthe lesionwasinsertedinto the cortex, the lesionwassegmentedascortex — thatis, thelesionappearedasa partialvolumeaveragingartifactandsowaslabelledascortex. Whenthe lesionwasplacedpartly in thecortex andpartly in thewhitematter, all voxelswith anunambiguouslesionintensityweresegmentedasle-sion,someof thevoxels(thosein thewhitematter)wereseg-mentedaslesionandsomeweresegmentedascortex. Thisis shown in Figure3. The segmentedtissueclassesare,inorderof decreasingintensity, white matter, white matterle-sion,grey matter, CSFandbackground.
Figure 4 shows a typical result from our white matter
lesionsegmentationmethodappliedto a patientwith mod-erate lesion burden. The image shows the effectivenessof our methodfor correctingthe misclassificationapparentwith statisticalclassification.Theimageshowsthatournewmethodimprovesthesegmentationof bothgrey matterandMS lesion.
Figure5 andFigure6 illustratethe segmentationof le-sionsdetectedautomaticallyfor 16 patientdatasets.Thesefiguresallow comparisonof the differencebetweenusingstatisticalclassificationandour new method.They demon-stratesthat partial volume averagingartifactsappearas alarge proportionof the voxels classifiedaslesionwith sta-tistical classificationalone.Figure7 is a magnifiedview ofonepair of theseimages. The segmentationwith the newmethodbetterdelineatesthewhite matterlesions,anddoesnot have the partial volumeaveragingerrorspresentin theEM segmentation.Thetwo largelesionson thepatient’s leftsideare larger anda clearpatternof periventricularlesioncanbe seen.This is obscuredin the EM segmentationbe-causethe EM segmentermisclassifiespart of theselesionsasgrey matter.
4 Discussionand Conclusion
Intensitybasedstatisticalclassificationhasbeenusedto seg-mentMRI scansinto differenttissueclasses.However, whentissueclassdistributions overlapsomevoxels are misclas-sified. This makesthe segmentationof MS lesionspartic-ularly difficult becausethe intensitydistribution of MS le-sionsstronglyoverlapsotherclasses(especiallygrey matter)in doubleechospinechoMRI.
We have developeda generalmethodfor automaticallyincorporatingknowledgeof normalbrain anatomyinto thesegmentationprocess. This allows correct segmentationof different structuresthat have a similar pixel intensityrange. Elastic matchingcan be usedto segmentsubcor-tical structuresdirectly. The segmentationachieved usingelasticmatchingis alsoableto distinguishbetweendifferentstructuresof thesametissueclass.Elasticmatchingallowsboundariesbetweenstructuresnotclearlyresolvedin MRI tobedeterminedusingthenormalanatomyasa model.How-ever, the elasticmatchingtechniqueis not ableto segmentstructureswhich arenot presentin the atlas,suchaswhitematterlesions.Theelasticdeformationmodelhasbeenableto matchsubcorticalanatomybut hasnot yet proven ableto modelthenormalanatomicalvariability exhibitedby thecortex.
Wehavedevelopedanew algorithmfor thesegmentation6
of thecortex. Table1 showsournew cortex segmentational-gorithmproducesa segmentationthatis consistentwith thatof manualsegmentation,andhasbetterreproducibilitysinceit producesthesamesegmentationeachtime it is repeated.We havedevelopeda methodfor theidentificationof there-gionof thewhitematterin MRI scansof thebrain.We haveshown that this allows for improvedsegmentationof whitematterlesions.
Statisticalclassificationexhibitstwo kindsof errorwhicharecorrectedby our new method.Thefirst is the incorrectclassificationof somevoxelsaswhite matterlesion,duetopartialvolumeaveragingandinherentclassintensitydistri-bution overlap.This error is correctedby relabellingvoxelsclassifiedas lesion if they areoutsidethe white matterre-gion. The secondis the incorrectclassificationof somele-sionvoxelsasgrey matter. Thiserroris correctedby reclas-sifying the white matterregion with a two classminimumdistanceclassifier.
We are investigatingthe applicationof this methodtothesegmentationof scansof patientswith braintumours.Inthesescansthe observed anatomycanbe significantlydis-tortedfrom thatof theatlas. We areinvestigatingtheauto-matic determinationof theanatomicallocationof eachMSlesionby referenceto theanatomicalatlas. We arealsoin-vestigatingtheuseof our segmentationmethodto studytheevolutionof MS lesionsover time.
2. DenglerJ, SchmidtM (1988)TheDynamicPyramid–A Model for Motion Analysiswith ControlledContinu-ity. InternationalJournalof PatternRecognitionandArtificial Intelligence2(2):275–286.
3. EttingerGJ,GrimsonWEL, Lozano-PerezT, Wells IIIWM, White SJ,Kikinis R (1994)AutomaticRegistra-tion for Multiple SclerosisChangeDetection.In IEEEWorkshoponBiomedicalImageAnalysis.
4. Filippi M, HorsfieldMA, Tofts PS,BarkhofF, Thomp-son AJ, Miller DH (1995) Quantitative assessmentof
15. Schmidt M, Dengler J (1989) Adapting Multi-GridMethods to the Class of Elliptic Partial DifferentialEquationAppearingin theEstimationof DisplacementVectorFields. In V Cantoni,R Creutzburg, S Levialdi,H Wolf (eds.), RecentIssuesin PatternAnalysisandRecognition,LectureNotesin ComputerScience399.Springer, Berlin-Heidelberg-New York-Tokyo pp.266–274.
7
LesionSegmentation
Bias Corrected Volume
Classified VolumeEM Segmentation
Elastic Match
DE SE MRI
Atlas Volume
Intracranial Cavity
SegmentationDeep Grey Matter Cortex
SegmentationWhite Matter
RegionWhite MatterLesions
Smooth Noise
Rigid Registration Matched Atlas
Figure1: Grey MatterandLesionSegmentationScheme
16. ShentonM, Kikinis R, McCarley R, SaiviroonpornP,Hokama H, Robatino A, Metcalf D, Wible C, Por-tas C, IosifescuD, Donnino R, GoldsteinJ, JoleszF(1995) Harvard Brain Atlas: A Teachingand Visual-izationTool. In M Loew, N Gershon(eds.), Proceed-ingsBiomedicalVisualization.IEEE ComputerSociety
Presspp.10–17.
17. Wells III WM, Grimson WEL, Kikinis R, JoleszFA(1994)Statisticalintensitycorrectionandsegmentationof MRI data. In SPIE Vol. 2359, Visualization inBiomedicalComputingpp.13–24.
8
(a)Sumof ratersegmentations (b) Sumof all segmentations
(c) Automaticsegmentation (d) Cortex from elasticmatch
(a)Away from cortex: EM segmenter (b) Away from cortex: New method
(c) Nearbycortex: EM segmenter (d) Nearbycortex: New method
(e) Insidecortex: EM segmenter (f) Insidecortex: New method
Figure3: Illustrationof lesionsegmentationwith EM segmenterandnew method.
10
(a) Earlyecho (b) Lateecho
(c) EM segmenter (d) New method
Figure4: Segmentationof white matter, grey matterandwhite matterlesion: The top row shows the early andlate echoimages,thebottomrow showsthesegmentationwith statisticalclassificationandwith ournew algorithm.
