GliomaTissueModelingbyCombingtheInformationofMRIandinvivoMultivoxelMRSWeibeiDOU,AoyanDONG,PingCHITsinghuaNationalLaboratoryforInformationScienceandTechnologyDept.ofElectronicEngineering,TsinghuaUniversity，Beijing,100084,P.R.Chinae-mail:ShaowuLINeuroimagingCenterofTiantanHospitalCapitalMedicalUniversity,Beijing,P.R.ChinaJean-MarcCONSTANSUnitdIRM,EA3916,CHRUCaen,FranceAbstractThispaperpresentsagliomamodelizationmethodandaregression-likemodeltocreateagraduallygliomaimage(GlioIm).Multimodalsignal,imagesofmagneticresonanceimaging(MRI)andinvivomultivoxelMRspectroscopy(MRS)arecombinedbytheregression-likemodelwithspatialresolutionregistration.ThismodelingmethodconsistsoffeaturemodelsofgliomasuchasthesignalintensityofMRimageandthemetabolitechangesofMRS,thecorrelationmodelnotedasmetabolitesratio(MetaR)andthecombinedregression-likemodel.TheestimatedGlioImincludesbothbrainstructureandgliomagradeinformation.Anonlinearmodelisproposedandvalidatedinthispaper.ThetestingdataisacquiredbySiemensTrioTim(3T)andSyngoMRB15atBeijingTiantanhospitalofChina.TheMRSofthreegliomapatients,twoaffectedbyastrocytomaandonebyglioma,andthechemicalshiftimaging(CSI)referenceT2imageswereconsideredinourvalidationexperiment.TheresultingGlioImsarecomparedwithgroundtruthprovidedbyneuroradiologistsofTiantanandverifiedwiththeirpathologyreport.Theyreportthatourmethodandmodelareveryefficient.Keywords-MRSpectroscopy；brain；glioma；chemicalshiftimaging；MRI；image；modeling；combinationI.INTRODUCTIONTodiagnosebraintissueabnormalities,liketumor,itsnecessarytousemultispectralmagneticresonanceimages(MRIs),suchasT1-weight,T2-weight,Gadolinium,FLAIRetcinordertofindsomeoftumorspropertiessuchassize,position,sort,andrelationshipwithothertissues,etc.Butthetumortypeandgradeareusuallydiagnosedfromhistopathologicalexaminationofasurgicalspecimen.However,Hydrogen1(1H)magneticresonancespectroscopy(MRS)isanon-invasiveMRtechniquethatprovidesbiochemicalinformationofmetabolites.Themajorbiochemicalcharacteristicscannoninvasivelyprovideusefulinformationonbraintumortypeandgrade1.Inmanystudies,invivo1H-MRShasbeenpresentedfordeterminingthetypeandgradeoftumors123.SinceinvivoMRSmeasurementsandanalysisaredependentontheacquisitiontechnicalthatcompromisethespatialresolutionandaccuracyforresultingmetabolitevalues4,metabolicchangeswithdiseaseisfrequentlysubtleanddiffuse.Furthermore,bychemical-shiftimaging(CSI)technique,themetaboliteimagesso-calledMRspectroscopicimaging(MRSI)canbecreatedbymultivoxelMRSinformation,butitisnotvisuallyinterpretableinthesenseofastructuralMRI4.Sothat,forthetumortissueclassification,itisimportantthatMRSIiscombinedwithMRItoestimatethevariationofmetabolitesandtoyieldmuchinformationregardingtissue.Duringmorethanadecade,automaticbraintumorclassificationbyMRShasbeendeveloped5,butthemorecleardefinitionofbraintumortypeandgrademaybeobtainedbycombinationofMRSIandMRI5.AtechniquetodifferentiateglioblastomafrommetastasislesionsbyusingMRIandMRSdatahasbeenpublishedin6.Wangetal.describedaclassificationofbraintumorsbyusingfeaturesselectionandfuzzyconnectednessin7,thesefeaturesareextractedfromMRIandMRSdata.TherearetwodifficultiesforcombingMRSIdataandMRIdata:firstly,thesedataarefromdifferentmodalities,sotheyarenotinthesamespatialresolution,verylowspatialresolutioninvoxelforMRSIandhighspatialresolutioninpixelforMRI.Secondly,oneMRimagecorrespondstothedistributionofalltissues,ortissuestructure.ButoneMRSimageisaprojectionimagewhichcorrespondstoonemetaboliteorratiobetweenseveralmetabolites.SothedifferentmetabolitevaluesmakevariationMRSimages,justlikethemappingofmetabolitedistributionsbyMRSIpresentedin8.ThequestionforapplicationishowtocombinetheseMRSimagesandMRimagestogiveanautomatictissueclassificationresult.ThekeypointofthecombinationishowtomodelthemetabolitedistributionfromMRS,whichcorrespondstoinformationfromMRimages.Forautomaticdescriptionofbraintumortypeandgrade,weproposeamodelizationmethodofgliomatissuesbycombingtheinformation,fromMRimagesandMulitivoxelMRSdata.ItcancreateaMRS-weightedMRimageautomaticallywhichkeepsthehighspatialresolutionlikeMRimageandthegreylevelscorrespondtothedeteriorationofbraintissues.ThesecondpartofthispaperintroducesthegliomatissuefeaturesbothinMRSvaluesandinMRimages.Thecombinationmodelingofthetwotypesofinformationispresentedinthethirdsectionanditsvalidationisshowninthefourthsection.Theconclusionaboutourresearchisgivenattheendofthispaper.ThisworkisfundedbyTsinghuaNationalLaboratoryforInformationScienceandTechnology（TNList）Cross-disciplineFoundation978-1-4244-4713-8/10/$25.002010IEEEII.FEATURESMODELOFGLIOMATISSUEFollowingtheresearchofdiagnosingbraintumorbyMRimagesandMRS,wecansummarizetwotypesofcharacteristicsofglioma,oneisthesignalintensityofT1-weightandT2-weightimages,andtheotheroneisthechemical-shiftvaluesofmetabolitespresentedbyMRSdata.A.SignalIntensityCharacteristicsofMRimagesWehaveproposedsomefuzzymodelingmethodsofdifferenttumorouscerebraltissuesonMRimagesbasedonfusionoftissuefeaturesin91011.TableIdescribesthecharacteristicsofbraintissuesbycreatingagradualityofsignalintensityasafunctionofdifferenttissuesandsequencesofMRI10,whereCSFistheabbreviationofcerebralspinalfluid,GMtheabbreviationofgraymatter,andWMwhitematter.IntableI,the“Seqs”isshortforSequencesofMRI”.Thesymbol“+”presentsahyper-signal；itmeansthatthesignalintensityisveryhighandtheimageisverybright.Thesymbol“-”presentsahypo-signal,theintensityisverylowandtheimageisverydark.Thesymbol“-+”meansthatthesignalintensityishigherthanhypo-signal,and“+-”meansthatitisdarkerthanhyper-signal.“-”meansthatthesignalintensityislowerthanthehypo-signal,and“+”meansthatitisbrighterthanthehyper-signal.AnexampleofT1-weightedimagenotedasT1,andT2-weightedimagenotedasT2areshowninFig.1TABLEI.SIGNALINTENSITYCHARACTERISTICSOFBRAINTISSUESONMRIMAGESSequencesGradualityofsignalintensityCSFGMWMGliomaEdemaNecrosisT1-+-+-T2+-+-(a)(b)Figure1.OriginalMRIimages(a)T1image,(b)T2imageB.MetaboliteChangesFeaturesofMRSTABLEII.SCALARDESCRIPTIONOFMETABOLITEVALUESMetabolitelevelabsentverylowlittlelowlowmediumlittlehighhighveryhighabbreviationAVLLLLMLHHVHThereareonlyseveralmetaboliteswhichcorrespondtogliomaamongalargenumberofmetabolitesofhumanbody.N-acetyl-asparate(NAA),creatine(Cr),choline(Cho),myo-inositol(mI),lactate(Lac)andfreelipids(Lip).ThevariationofthesemetabolitescanbeorderedinascalarformasshowninTableII,wherethescalarorderis:absent,verylow,littlelow,low,medium,littlehigh,high,veryhigh,whichcorrespondtometabolitevaluesfrom0tomaximum.ThemetabolicchangeswithbraintissuesareshowninTableIII.Itisconcludedfrom121314.TABLEIII.METABOLITECHANGESFEATURESOFBRAINTISSUESONMRSMetabolitevariationofmetabolitescorrespondingwithbraintissuesCSFGMWMGliomaEdemaNecrosisNAAVLVHHL/VLMAChoAMLHH/VHLHACrLHHM/LLLAmILMLHHLH/MALipAVLLHLVHLacLHVLAH/LHLHHIII.MODELIZATIONBYCOMBININGMRSWITHMRITheaimofthismodelizationstudyistocreateagradually?gliomaimage,notedasGlioIm,whichincludesbrainstructureandgliomagradeinformation.IfthegliomagradeinformationisconsideredasacorrelationfunctionbetweenMRsignalandpathologicalchanges.Weproposearegression-likemodeltoestimatetheGlioImfromMRimagesnotedasMRImandmetabolitechanges.A.CorrelationmodelOneofthecorrelationfunctionsismetabolitechangescorrespondingtoglioma.BycombiningtheinformationinTableIandTableIII,wecanrebuildaconclusionTableIVaboutgliomacharacteristicswithrelativequantizationofmetabolitesofTableIII.Therelativequantizationisratiosbetweenmetabolitevalues,suchastheratioofChoandNAAnotedasCho/NAAinTableIV,itiscalledmetabolitesratio(MetaR),andTableIViscalledcorrelationmodelinthispaper.TABLEIV.METABOLITESRATIOCHARACTERISTICOFBRAINTISSUESMetabolitevariationofmetabolitescorrespondingwithbraintissuesCSFGMWMGliomaEdemaNecrosisCho/NAAAVLLVHHACho/CrALLHHAmI/CrMLMHHALip/CrAVLVLHMVHLac/CrLHVLAHHHTheMetaRcharacteristicsofglioma,edemaandnecrosisareenhancedandthenormaltissuesarereduced.TheyassortwithsignalintensitycharacteristicsofT2-weightedimagedescribedinTableI.B.Regression-likemodelwithspatialresolutionregistrationNormaly,MetaRisafunctionofvoxeldecidedbyCSIsliceshowninFig.2.Sothat,itisatwodimensionalfunctionnotedasMetaR(i,v),where“i”isindexofmetaboliteand“v”istheindexofvoxelcorrespondedwithCSIslice.Asthesamereason,GlioImcanbecreatedasathreedimensionalfunction,notedasGlioIm(v,p,g),where“p”isindexofpixelcorrespondedwithMRIm,and“g”isthegreylevelofselectedMRimageandcorrespondsto“p”.Infact,MRImisatwodimensionalfunctionnotedasMRIm(p,g),whereandgG，1,2,.TTPDFLAIRGadoDiffusionPerfusion=GConsidertwovariables,MRImandGlioIm,MRImisacertainimagelikeT2,GlioImisanestimatedimage.ThecorrelationmodelMetaRcanbeconsideredasonerelationshipbetweenthem.Sotheregression-likemodelforestimatingGlioImfromMRImcanbecreatedasequation(1).Im(,)(,)Im(,)GliovpgMetaRivMRpg=(1)Where“”notesanecessaryoperator,and“p”correspondsto“v”.Ifalinearregressiveisnecessary,equation(1)canberewrittenas(2):Im(,)(,)Im(,)(,)GliovpgMetaRivMRpgMetaRjv=+(2)where“i”and“j”indicatedifferentmetabolites.C.NonlinearRegression-likemodelToavoidmosaiceffects,weproposeanonlinearregression-likemodelwithspatialresolutionregistrationin(3).Im(,)Im(,)exp(,)(.)MRpgGliovpgMetaRivMetaRjvT=+(3)where“T”isatimeconstantcorrespondingtoMRIm(p,g).AccordingtothecorrelationmodelofTableIV,theLip/CrandLac/Crarespecificfeatureswhicharedependentonthetumorgrade.Sothat,inthemodelofequation(2),wehave:,/,/,/,/ijMetaRChoNaaChoCrmICrLipCrLacCr=IJIJIJ，BecausetheJofMetaRisthegrademarker,ittakesaninterceptiveroletomakeadifferentgreylevelfromothervoxelsandindicatesavariablegrade.IV.VALIDATIONANDRESULTA.MaterielThreegliomapatients,twoaffectedbyastrocytomaandonebyglioma,wereconsideredinourvalidationexperiment.ThetestingdataareadatapairconsistedofCSIrawdataandtheirreferenceimages.ThesedatawereacquiredwithSTEAMsequenceatBeijingTiantanhospital(China),bySiemensMRTrioTim(3T)andsyngoMRB15.TheMRSrawdataaremeasuredbycsi_st/90protocolwithTR3000/TE72/TM6.T2-weightedimagesaremeasuredbyt2_tse_traprotocolwithTR4500/TE80.TwoexamplesofthesedataareshowninFig.2.Thenonlinearregression-likemodel(3)isvalidatedbyourtestingexperimentation.MRImof(3)isT2with0.570.57mm2pixelsizeand5mmslicethickness.ThetimeconstantTin(3)isindicatedbyhistogrampeakofCSIreferenceimagesinT2.ThemetabolitevaluesarecalculatedbyTHU-MRSv0.5developedbyourresearchgroupandpublishedin15.TheCSIslicesnotethattheMRSvoxelsizeis141420mm3.(a)(b)Figure2.ExampleofCSIslice(down-left)withitsreferenceimagesandmetabolitesvaluescorrespondedwithvoxelsize141420mm3.(a)fromanastrocytomapatient,masculine30yearsold.(b)fromagliomapatient,feminie48yearsold.B.ResultThevalidationresultscorrespondedtoVOIareshowninFig.3(f)andFig.4(f).Thehighersignalorbrighterpixelin(f)marksgreaterpossibilityofgliomaorhighertumorgrade.InFig.3and4,(a)aretheoriginalT2-weightedimageswiththesignofVOI,(b)arethehandlabelresultsas“Groundtruth”fromneuroradiologistsofTiantan,(c)areonepartof(a)inVOI,(d)aretheresultsofexponentialcomponentofequation(3)whichpresentsthecombinedinformationofT2andCho/Naa,(e)aretheresultsofsuperpositionofT2and(Lip+Lc)/Cr.(a)T2+VOI(b)Groundtruth(c)OriginalT2inVOI(d)MetaR(Cho/Naa)(e)MetaR(Lac+Lip)/Cr)(f)ResultingGlioImFigure3.ResultingGlioIm(f)ofthepatientaffectedbyastrocytomaC.DiscussionThebrighterpixelinFig.3(d)or(f)denotesnotonlyhigherCho/NaabutalsobrighterT2.BecauseMetaRvaluesinTableIVareconsistentwiththeintensityofT2.Soitmayindicategliomaandhighergraderegion.ThedarkerpixelspresentlowerCho/NaaanddarkerT2,mayindicatenormaltissues.Thentherearesomebrighterpixelsin(d)and(f),theyarenotconnectedwithgliomaregion,theyareCSFperhaps,becauseCSFisbrighterinT2.WecanremovethemsimplybyusingregisteredFLAIRimage.TheregisteredGadoliniumimagealsocanbeusedtoindicateenhancedpixelsorregion.(a)T2+VOI(b)Groundtruth(c)originalT2inVOI(d)MetaR(Cho/Naa)(e)MetaR(Lac+Lip)/Cr)(f)ResultingGlioImFigure4.ResultingGlioIm(f)ofthepatientaffectedbygliomaBecauseamongthe5metaboliteratiosinTableIV,onlythreepresentevidentchanges,likeCho/Naa,Lip/CrandLac/Cr.Theothertworatiosarenotutilizedinourexperiment.ItispossibletouseotherMRIsequencessuchasT1,butitisnecessaryeithertotransformgreylevelsofimageortoinversethevalueofMetaR.Asmentionedin16,animageresultedfromfusionofgliomafeaturesextractedfrommultimodalitysignal,aspresentedin9,canalsobeusedasMRIminthisgliomamodel.V.CONCLUSIONAdvantagesofMRItechniqueprovidemorepossibilitywithmulti-sequencesandmultimodalitiessignaltorealizethetumordiagnosis,treatmentandprognosis.Butitisheavyworkforprocessingallsignalstodoafinaldecision.SoAutomaticquantificationandcombinationanalysisisveryimportantandthemodelingoftumorfeaturesisthekeypointforperformingit.Wehaveproposedaframeworkoffuzzyfeaturesfusionsystemin16andpublishedsomeresearchresultsaboutfusingthetumorfeaturesextractedfromT1,T2andprotondensityimages9.Inthispaper,wepresenttheprimarystudyaboutthetumorfeaturescombinationofMRSandMRimages.Theproposedmodelingmethodandnonlinearregression-likemodelarevalidforseparatingthebraintissuesespeciallyglioma.Itwillbeusedfortumortissuesclassification,segmentation,tumortypeandgradedecision,etc.Thereisstillmuchworktoimprovethismodelandtointegrateitwiththefusionsysteminthefuture.REFERENCES1HoweFA,BartonSJ,CudlipSA,StubbsM,SaundersDE,MurphyM,WilkinsP,OpstadKS,DoyleVL,McLeanMA,BellBA,GriffithsJR.“Metabolicprofilesofhumanbraintumorsusingquantitativeinvivo1Hmagneticresonancespectroscopy”.MagnResonMed.2003Feb；49(2):223-32.2PreulMC,CaramanosZ,CollinsDL,VillemureJ-G,LeblancR,OlivierA,PokrupaR,ArnoldD.Accurate,non-invasivediagnosisofhumanbraintumorsbyusingprotonmagneticresonancespectroscopy.NatMed1996；2:323325.3MajsC,AguileraC,CosM,CaminsA,CandiotaAP,Delgado-GoiT,SamitierA,CastaerS,SnchezJJ,MatoD,AcebesJJ,ArsC.“Invivoprotonmagneticresonancespectroscopyofintraventriculartumoursofthebrain”EurRadiol.2009Aug；19(8):2049-59.4A.A.Maudsley,C.Domenig,V.Govind,A.Darkazanli,C.Studholme,K.Arheart,C.Bloomer,“MappingofbrainmetabolitedistributionsbyvolumetricprotonMRspectroscopicimaging(MRSI)”MagneticResonnanceinMedicin61:548-559(2009).5Garcia-GomezJ.,LutsJ.,Julia-SapeM.,KrooshofP.,TortajadaS.,VicenteJ.,MelssenW.,Fuster-GarciaE.,OlierI.,PostmaG.,MonleonD.,Moreno-TorresA.,PujolJ.,CandiotaA.-P.,Martinez-BisbalM.C.,SuykensJ.A.K.,BuydensL.,CeldaB.,VanHuffelS.,ArusC.,RoblesM.,Multiproject-multicenterevaluationofautomaticbraintumorclassificationbymagneticresonancespectroscopy,MagneticResonanceMaterialsinPhysics,BiologyandMedicine,vol.22,Feb.2009,pp.5-18.6LutsJ.,LaudadioT.,Martinez-BisbalM.C.,VanCauterS.,MollaE.,PiquerJ.,SuykensJ.A.K.,HimmelreichU.,CeldaB.,VanHuffelS.,DifferentiationbetweenbrainmetastasesandglioblastomamultiformebasedonMRI,MRSandMRSI,inProc.ofthe22ndIEEEInternationalSymposiumonComputer-BasedMedicalSystems(CBMS),Albuquerque,NewMexico,Aug.2009,pp.1-8.7QiangWang,EiriniKaramaniLiacouras,EricksonMiranda,UdayS.Kanamalla,andVasileiosMegalooikonomou,ClassificationofbraintumorsusingMRIandMRSdata,Proc.SPIE6514,(2007)pp.65140S-18.8A.A.Maudsley,C.Domenig,V.Govind,A.Darkazanli,C.Studhol