Estimating CT image from MRI data using stru.docx
- 文档编号:9190350
- 上传时间:2023-05-17
- 格式:DOCX
- 页数:29
- 大小:139.91KB
Estimating CT image from MRI data using stru.docx
《Estimating CT image from MRI data using stru.docx》由会员分享,可在线阅读,更多相关《Estimating CT image from MRI data using stru.docx(29页珍藏版)》请在冰点文库上搜索。
EstimatingCTimagefromMRIdatausingstru
IEEETMI-2015-03061
EstimatingCTImagefromMRIDataUsingStructuredRandom
ForestandAuto-contextModel
TriHuynh、YaozongGao卞,JiayinKang,LiWang,PeiZhang,JunLian,DinggangShen*,Senor
Member,IEEEandfortheAlzheimer5sDiseaseNeuroimagingInitiative(ADNI)
Fig.1.ApairofcorrespondingMRandCTimagesfromthesamehumanbrain.BothairandbonehaveverylowresponsesinMRimage,buttheyarehighlydifferentiableinCTimage.
Abstract—Computedtomography(CT)imagingisanessentialtoolinvariousclinicaldiagnosesandradiotherapytreatmentplanning.SinceCTimageintensitiesaredirectlyrelatedtopositronemissiontomography(PET)attenuationcoefficients,theyareindispensableforattenuationcorrection(AC)ofthePETimages.However,duetotherelativelyhighdoseofradiationexposureinCTscan,itisadvisedtolimittheacquisitionofCTimages.Inaddition,inthenewPETandmagneticresonance(MR)imagingscanner,onlyMRimagesareavailable,whichareunfortunatelynotdirectlyapplicabletoAC.TheseissuesgreatlymotivatethedevelopmentofmethodsforreliableestimateofCTimagefromitscorrespondingMRimageofthesamesubject.Inthispaper,weproposealearning-basedmethodtotacklethischallengingproblem.Specifically,wefirstpartitionagivenMRimageintoasetofpatches.Then,foreachpatch,weusethestructuredrandomforesttodirectlypredictaCTpatchasastructuredoutput,whereanewensemblemodelisalsousedtoensuretherobustprediction.Imagefeaturesareinnovativelycraftedtoachievemulti-levelsensitivity,withspatialinformationintegratedthroughonlyrigid-bodyalignmenttohelpavoidingtheerror-proneinter-subjectdeformableregistration.Moreover,weuseanauto-contextmodeltoiterativelyrefinetheprediction.Finally,wecombineallofthepredictedCTpatchestoobtainthefinalpredictionforthegivenMRimage.Wedemonstratetheefficacyofourmethodontwodatasets:
humanbrainandprostateimages.Experimentalresultsshowthatourmethodcan
Copyright(c)2010IEEE.Personaluseofthismaterialispermitted.However,permissiontousethismaterialforanyotherpurposesmustbeobtainedfromtheIEEEbysendingarequesttopubs-permissions@ieee.org.
rTriHuynhandYaozongGaoareco-firstauthors.
*DinggangShenisthecorrespondingauthor.
PartofthedatacollectionandsharingforthisprojectwasfundedbytheAlzheimer'sDiseaseNeuroimagingInitiative(ADNI)(NationalInstitutesofHealthGrantU01AG024904)andDODADNI(DepartmentofDefenseawardnumberW81XWH-12-2-00l2).ThisworkwaspartiallysupportedbyNIHgrants(EB006733,EB008374,EB009634,MH100217,AG041721,
AG042599,CA140413).
T.HuynhiswiththeIDEAlab,DepartmentofRadiologyandBRIC,UniversityofNorthCarolinaatChapelHill,NC,USA.(email:
hquoctri@)
Y.GaoiswiththeDepartmentofComputerScience,andalsowiththeIDEAlab,DepartmentofRadiologyandBRIC,UniversityofNorthCarolinaatChapelHill,NC,USA.(email:
yzgao@cs.unc.edu)
J.Kang,L.Wang,andP.ZhangarewiththeIDEAlab,DepartmentofRadiologyandBRIC,UniversityofNorthCarolinaatChapelHill,NC,USA.
J.LianiswiththeDepartmentofRadiationOncology,UniversityofNorthCarolinaatChapelHill,NC,USA.(email:
jun_lian@med.unc.edu)
*D.SheniswiththeBiomedicalResearchImagingCenter,UniversityofNorthCarolinaatChapelHill,NC27599,USA,andalsowiththeDepartmentofBrainandCognitiveEngineering,KoreaUniversity,Seoul136-071,Korea(email:
dgshen@med.unc.edu)
accuratelypredictCTimagesinvariousscenarios,evenfortheimagesundergoinglargeshapevariation,andalsooutperformstwostate-of-the-artmethods.
IndexTerms—CTPrediction,PETAttenuationCorrection,DosePlanning,RandomForest,Auto-context
I.INTRODUCTION
Computedtomography(CT)imagingiswidelyusedinvariousmedicalpractices,e.g.,fordetectinginfarction,tumors,calcifications,etc.Besides,CTimagesarealsoindispensableforattenuationcorrection(AC)ofpositronemissiontomography(PET)images[1]inthePET/CTsystem.ACisaprocessofderivingattenuationmapfromtheCTimagetocorrectthecorrespondingPETimage,basedontheradiation-attenuationpropertiesoftissuesrevealedbytheattenuationmap.
However,theuseofCTscanisadvisedtobelimited,consideringtheriskofradiationexposure.Forexample,ithasbeenshownthatasmanyas0.4%ofcancersintheUSareduetoCTscanningperformedinthepast,andthisnumbermayincreasetoashighas1.5to2%inthefuture[2].Besides,asanemergingimagingtool,therecentPETandmagneticresonanceimaging(MRI)systemreplacesCTbyMRI.However,itisintrinsicallyhardtopredictattenuationcoefficientsfromMRimages,sincetheMRIsignalsofindividualvoxelsarerelatedtoprotondensity,nottheelectrondensityinformationthatisrequiredforAC.Forexample,usingmoststandardMRIsequences,bothairandcompactbonegenerateverylowsignals,whereastheirattenuationcoefficientsarehighlydifferent,asshowninFig.1.
contrast
IEEETMI-2015-03062
Hence,thereisacrucialneedforpredictingaCTimagefromanMRimage.Theexistingworkscanberoughlyclassifiedintothefollowingfourcategories:
•Tissuesegmentationbasedmethods.ThebasicideaistofirstsegmentanMRimageintodifferenttissueclasses,andthenassigneachclasswithaknownattenuationproperty.However,thistypeofmethodsmayfailduetotheexistenceofsomeambiguoustissueclasses,suchasairandbone.Inparticular,Zaidietal.[3]proposedafuzzylogicbasedmethodtosegmentMRIintofivetissueclasses,andmanuallyfixedfailuresofautomaticsegmentation.Hsuetal.[4]performedsegmentationusingmultipleMRImodalities,aswellasconsideringthefatandwaterimagevolumeswithDixonMRIsequence.TheyfoundthatasingleMRIvolumeisinsufficienttoseparatealltissueclasses.Similarly,Berkeretal[5]combinedUTE/DixonMRIsequencetosegmenttissuesforAC.
•Atlas-basedmethods.Thesemethodsestimatetheattenuationmapofagivensubjectbywarpingtheattenuationmapofanatlastothissubject[6,7],usingthedeformationfieldestimatedbyregistrationoftheMRimagesbetweentheatlasandthesubject.However,theperformanceofthesemethodsishighlydependentontheregistrationaccuracy.
•Learning-basedmethods.TheMR-CTrelationshipcanbelearnedfromatrainingsetandthenappliedtoatargetMRimageforCTimageprediction.Sinceitisnoteasytolearnsuchrelationshipfromasinglemodality,Johanssonetal.[8]proposedtobuildaGaussianmixtureregressionmodelforlearningfrommultiplemodalities,i.e.,twoUTEimagesandoneT2-weightedMRimage.TheideaofusingGaussianmixturemodelwasalsoadoptedbyRoyetal.[9]tosynthesizeCTimagefromtwoUTEMRimages.ThesametechniquewasalsoappliedtoestimateCTimagefromoneMRimage[10].However,thespatialinformationwasdisregardedandthenthequalityofestimatedCTimagewasmodest,thusoftenusedastheintermediatemeanstofacilitatethesubsequentregistrationtask.
•Integrationofatlas-basedandpatternrecognitionmethods.Afterwarpingatlasestothetargetimage,alocalregressionmodel(characterizedbybothspatiallocationsandimagepatchintensity)canbebuiltandappliedtothetargetimage.Forexample,Hofmannetal.[7]usedaGaussianprocessregressionmodel.Althoughthistypeoftechniquescanproducepromisingresults,theirperformancestillhighlydependsontheaccuracyofthedeformableregistrationbetweentheatlasandtargetMRimages.
Anothercloselyrelatedresearchfieldisimagesynthesis,whichhasasimilargoalofsynthesizingoneimagemodalityfromothermodalities,althoughwithdifferentapplications.Mostofsuchstudiesfallunderthefollowingtwocategories:
•Learning-basedmethods.Jogetal.usedrandomforesttoreconstructthehigh-resolutionT2-weightedMRimagefrombothlow-resolutionT2-weightedandhigh-resolutionT1-weightedMRimages[11].AsimilartechniquewasalsousedtoestimateFluidAttenuatedInversionRecovery(FLAIR)sequencefromT1,T2,andPD-weightedMRsequences[12].Therandomforestsusedintheseworksarerathergeneralandsimple,whichneitherusedspatialinformationnorhadspecificenhancementsforCTimageprediction.In[13],Lietal.usedtheconvolutionalneuralnetworkstoestimatePETimagefromMRimage.However,neuralnetworkapproachsuffersfromlongtrainingtime,rangingfromdaystoweeks,anditsperformancehighlydependsonthesuccessfultuningofmanyparameters.
•Examplar-basedmethods.Adominantlineofresearchinthiscategoryisthesparserepresentationbasedmethods.First,thetargetimagepatchissparselyrepresentedbyasetofatlaspatchesofthesamemodality.Then,theresultingsparsecoefficientsareusedtointegratethecorrespondingatlaspatchesofanothermodalitytoestimatethedesiredimagepatchinthatmodalityforthetargetsubject.Royetal.usedthistechniquetosolvemultipleproblems,i.e.,predictingmagnetizationpreparedrapidgradientechosequencefromspoiledgradientrecalledsequenceandviceversa[14,15],aswellaspredictingFLAIRimagefromT1-andT2-weightedMRI[16].Yeetal.[17]estimatedT2-anddiffusion-weightedimagesfromT1-weightedMRI,whileIglesiasetal.[18]predictedTi-weightedMRIfromPD-weightedMRI.Butonemaindrawbackofthesemethodsisthatthepredictionisoftenverycomputationallyexpensiveduetohugeoptimizationneededinthetestingphase.
Inthispaper,wewillemployanenhancedrandomforestto
specificallytackletheproblemofestimatingCTimagefrom
MR
- 配套讲稿:
如PPT文件的首页显示word图标,表示该PPT已包含配套word讲稿。双击word图标可打开word文档。
- 特殊限制:
部分文档作品中含有的国旗、国徽等图片,仅作为作品整体效果示例展示,禁止商用。设计者仅对作品中独创性部分享有著作权。
- 关 键 词:
- Estimating CT image from MRI data using stru