醫(yī)療影像分割挑戰(zhàn)賽方案分享：14億參數(shù)醫(yī)療分割大模型

DoespretrainedSTU-NetworkinMICCAIChallenge?2StatisticsStatisticsofMICCAIChallengesCNNCNNorTransformer?4ComparisonofCNNandTransformer567LargeLargedatasetinmedicalimagingTotalSegmentator(20。9STUSTU-NetSTUSTU-NetSTUSTU-NetonMICCAI2023ChallengesSTUSTU-NetonMICCAI2023ChallengesTask1:AdultGliomawith1251dataTask2:GliomainAfricawith60dataTask3:Meningiomawith1000dataTask4:BrainMetastaseswith238dataTask5:PediatricTumorswith99dataSTUSTU-NetonMICCAI2023ChallengesAbdomenCTAAorticVesselTree(AVT)OfflineexperimentsOfflineexperimentsATLAS2023SettingSettingforsubmissionATLAS2023222--?Largerpatchsize:TrainingorTesting?Pretrainmodel?Externaldata?Mixtrainingofdifferentmodalities?Largerpatchsize:TrainingorTesting?Pretrainmodel?Externaldata?MixtrainingofdifferentmodalitiesTaskPatchSizeAvg.BraTS23GLI0.78760.82210.83660.8154BraTS23GLI0.77840.84910.8470.82480.8248BraTS23SSA0.82650.84640.5690.7473BraTS23SSA0.74960.79420.80660.78350.78350.85450.86050.84020.85170.85550.86080.84610.85410.8541----0.62580.66830.60220.63210.46920.73380.81180.67160.46980.79050.84440.70160.7016DetailsforFLAREChallengeIntroductionIntroductionofFLAREChallenge?Fully-supervisedsetting:?DonotconsiderEfficiency?UseDSCandNSDasmetrics1.Howtouse2000unlabeledimagesinnnU-Netframework?2.HowtomakennU-Netefficient?PseudoLabelingnnU-NetframeworkEfficiencyAccuracyInputpatchsizeEnsembleInputresolutionInferencestepsizeModelsizeTesttimeaugmentation?WecangeneratehighqualitypseudolabelsforunlabeledimagesinnnU-Net,butnotefficient??WeCanwemakennU-Netinferencefastandlow-resourcecomsumption,butnotaccurate?pseudolabelsbignnUNetunlabeledimagespseudolabelsbignnUNetunlabeledimagesbignnUNetpseudopseudolabelsComparisonofourtwonnU-Net2325541.Train5bignnUNetmodelsby5-foldcross-validationonthelabeleddata2.Predictone-hothardpseudolabelsonunlabeleddatausingourdesignedaccurateinferencesettingwith5-foldensembleofbignnUNet.3.Iterativere-trainabignnUNetontheunionoflabeleddataandunlabeleddatawithpseudolabelsandthengeneratenewone-hothardpseudolabelsforthenextround.(repeat)4.Selectpseudolabelsbasedonstabilityofpseudolabelsduringdifferenttrainingrounds.5.TrainasmallnnUNetontheunionoflabeleddataandselectedunlabeleddatawithpseudolabelsforfinalevaluation.SlidingWindowStrat?LimitedVolume?Locateinthemiddleineachxyplane?Spendtoomuchtimeonthebackgroundarea(4mm,1.2mm,1.2mm)×(32,128,192)×(1+0.5×2)=(256mm,307.2mm,460.8mm)EnoughtocovertheabdominalregionineachxyplaneIfthemiddlewindowhasnoforegroundprediction,wecanskipthesurroundingwindowsCodeoptimizationforefficien1.Deletethecropoperation2.Changetheresamplefunction(skimage->torch)3.Conditional“all_in_gpu”(forsmallimage)?BignnUNetachievesmeanDSCof0.9064andranksfirstononlineleaderboard?Trainingwithpseudolabelscangreatlyimprovemodelsperformance?Theimprovementboostfrompseudolabelsismoresignificantforsmallmodel Testenvironment:?Amodelinitiallytrainedonfully-labeledcasestoiterativelygeneratepseudolabelsforunlabeledcases?nnU-Netv1basedinferenceacceleration?Nofully-labeledcase?nnU-Netv2basedaccelerationisexpected(2)Phase2:Pseudo-Casespacing:(1.5×0.875×0.875)skimage.transform.resize->erpolateresampledlogits->nonlinear->argmax(original)argmax->resampleseg(optimized)4.AvoidTestTimeAugmentation(TTA)5.EfficientslidingAblationstudywithin20seconds.Additionally,theGPUmemoryconsumptionstAwardsAwardsThirdplaceonBraTS-Africa2023SecondplaceonBraTS-Men2023ThirdplaceonBraTS-Met2023A-Eval(Abdomen-EvalA-Eval(Abdomen-Eval)OtherworksofourgroupTraditionalTraditionalEvaluationSAMSAM-Med2D&SA-Med2D-20MSAM-Med3DSAM-Med3D?葉錦(JinYe)–莫那什大學?李添試(TianbinLi)–電子科技大學?蘇炎洲(YanzhouSu)–電子科技大學?程俊龍