Deep Network Design for Medical Image Computing Principles and Applications THE ELSEVIER AND MICCAI SOCIETY BOOK SERIES Advisory Board NicholasAyache JamesS.Duncan AlexFrangi HayitGreenspan PierreJannin AnneMartel XavierPennec TerryPeters DanielRueckert MilanSonka JayTian S.KevinZhou Titles Balocco,A.,etal.,ComputingandVisualizationforIntravascularImagingand ComputerAssistedStenting,9780128110188. Dalca,A.V.,etal.,ImagingGenetics,9780128139684. Depeursinge,A.,etal.,BiomedicalTextureAnalysis,9780128121337. Munsell,B.,etal.,Connectomics,9780128138380. Pennec,X.,etal.,RiemannianGeometricStatisticsinMedical ImageAnalysis,9780128147252. Trucco,E.,etal.,ComputationalRetinalImageAnalysis,9780081028162. Wu,G.,andSabuncu,M.,MachineLearningandMedicalImaging,9780128040768. ZhouS.K.,MedicalImageRecognition,SegmentationandParsing,9780128025819. Zhou,S.K.,etal.,DeepLearningforMedicalImageAnalysis,9780128104088. Zhou,S.K.,etal.,HandbookofMedicalImageComputingandComputer AssistedIntervention,9780128161760. Deep Network Design for Medical Image Computing Principles and Applications Haofu Liao S. Kevin Zhou Jiebo Luo AcademicPressisanimprintofElsevier 125LondonWall,LondonEC2Y5AS,UnitedKingdom 525BStreet,Suite1650,SanDiego,CA92101,UnitedStates 50HampshireStreet,5thFloor,Cambridge,MA02139,UnitedStates TheBoulevard,LangfordLane,Kidlington,OxfordOX51GB,UnitedKingdom Copyright©2023ElsevierInc.Allrightsreserved. Nopartofthispublicationmaybereproducedortransmittedinanyformorbyanymeans,electronic ormechanical,includingphotocopying,recording,oranyinformationstorageandretrievalsystem, withoutpermissioninwritingfromthepublisher.Detailsonhowtoseekpermission,further informationaboutthePublisher’spermissionspoliciesandourarrangementswithorganizationssuch astheCopyrightClearanceCenterandtheCopyrightLicensingAgency,canbefoundatourwebsite: www.elsevier.com/permissions. Thisbookandtheindividualcontributionscontainedinitareprotectedundercopyrightbythe Publisher(otherthanasmaybenotedherein). Notices Knowledgeandbestpracticeinthisfieldareconstantlychanging.Asnewresearchandexperience broadenourunderstanding,changesinresearchmethods,professionalpractices,ormedicaltreatment maybecomenecessary. Practitionersandresearchersmustalwaysrelyontheirownexperienceandknowledgeinevaluating andusinganyinformation,methods,compounds,orexperimentsdescribedherein.Inusingsuch informationormethodstheyshouldbemindfuloftheirownsafetyandthesafetyofothers,including partiesforwhomtheyhaveaprofessionalresponsibility. Tothefullestextentofthelaw,neitherthePublishernortheauthors,contributors,oreditors,assume anyliabilityforanyinjuryand/ordamagetopersonsorpropertyasamatterofproductsliability, negligenceorotherwise,orfromanyuseoroperationofanymethods,products,instructions,orideas containedinthematerialherein. ISBN:978-0-12-824383-1 ForinformationonallAcademicPresspublications visitourwebsiteathttps://www.elsevier.com/books-and-journals Publisher:MaraE.Conner AcquisitionsEditor:TimPitts EditorialProjectManager:EmilyThomson ProductionProjectManager:SuryaNarayanan Jayachandran CoverDesigner:ChristianJ.Bilbow TypesetbyVTeX Toourfamiliesfortheirenduringsupport This page intentionally left blank Contents Listoffigures ................................................. xi Acknowledgments ............................................. xvii CHAPTER 1 Introduction .................................... 1 1.1 Medicalimagecomputing ........................... 1 1.1.1 Medicalimagereconstruction ................... 2 1.1.2 Medicalimageanalysis ........................ 2 1.1.3 Medicalimagecomputingasfunctionalapproximation 3 1.2 Deeplearningdesignprinciples ....................... 4 1.2.1 Computervisiontechniquesformedicalimage computing.................................. 4 1.2.2 Machinelearningtechniquesformedicalimage computing.................................. 5 1.2.3 Medicaldomainknowledge..................... 5 1.3 Chapterorganization ............................... 6 References....................................... 8 CHAPTER 2 Deep learning basics ........................... 11 2.1 Convolutionalneuralnetworks ........................ 11 2.1.1 3Dconvolutionalneuralnetworks ................ 12 2.2 Recurrentneuralnetworks ........................... 13 2.2.1 Longshort-termmemory....................... 14 2.2.2 BidirectionalRNN ........................... 15 2.3 Deepimage-to-imagenetworks ....................... 16 2.3.1 Retainingspatialresolutions .................... 16 2.3.2 Fullyconvolutionalnetworks.................... 17 2.3.3 Encoder–decodernetworks ..................... 17 2.4 Deepgenerativenetworks ........................... 18 2.4.1 Basicmodels................................ 19 References....................................... 21 PART 1 Deep network design for medical image analysis and selected applications CHAPTER 3 Classification: lesion and disease recognition .... 27 3.1 Designprinciples .................................. 28 3.1.1 Choiceofdeepneuralnetworks .................. 28 3.1.2 Choiceofclassificationtasksandobjectives ........ 29 3.1.3 Transferlearning ............................. 31 3.1.4 Multitasklearning ............................ 32 vii viii Contents 3.2 Casestudy:skindiseaseclassificationversusskinlesion characterization ................................... 33 3.2.1 Background ................................ 35 3.2.2 Dataset .................................... 35 3.2.3 Methodology ............................... 37 3.2.4 Experiments ................................ 38 3.2.5 Discussion ................................. 42 3.3 Casestudy:skinlesionclassificationwithmultitasklearning . 44 3.3.1 Background ................................ 44 3.3.2 Dataset .................................... 45 3.3.3 Methodology ............................... 47 3.3.4 Experiments ................................ 49 3.3.5 Discussion ................................. 52 3.4 Summary ........................................ 53 References....................................... 54 CHAPTER 4 Detection: vertebrae localization and identification 59 4.1 Designprinciples .................................. 60 4.1.1 Choiceofdeepneuralnetworks .................. 60 4.1.2 Choiceofdetectiontasksandobjectives ........... 64 4.2 Casestudy:vertebraelocalizationandidentification ........ 67 4.2.1 Background ................................ 69 4.2.2 Methodology ............................... 70 4.2.3 Experiments ................................ 76 4.2.4 Discussion ................................. 80 4.3 Summary ........................................ 81 References....................................... 83 CHAPTER 5 Segmentation: intracardiac echocardiography contouring ..................................... 89 5.1 Designprinciples .................................. 90 5.1.1 Choiceofdeepneuralnetworks .................. 90 5.1.2 Choiceofsegmentationtasksandobjectives ........ 92 5.1.3 Imagerestorationforsegmentation ............... 94 5.2 Casestudy:intracardiacechocardiographycontouring ...... 96 5.2.1 Methodology ............................... 97 5.2.2 Experiments ................................ 99 5.2.3 Discussion ................................. 102 5.3 Summary ........................................ 102 References....................................... 103 CHAPTER 6 Registration: 2D/3D rigid registration............. 109 6.1 Designprinciples .................................. 111 6.1.1 Deepsimilaritybasedregistration ................ 111 6.1.2 Reinforcementlearningbasedregistration .......... 112 Contents ix 6.1.3 Supervisedtransformationestimation ............. 113 6.1.4 Unsupervisedtransformationestimation ........... 115 6.2 Casestudy:2D/3Dmedicalimageregistration ............ 117 6.2.1 Problemformulation .......................... 119 6.2.2 Methodology ............................... 121 6.2.3 Experiments ................................ 125 6.2.4 Limitations ................................. 130 6.2.5 Discussion ................................. 130 6.3 Summary ........................................ 130 References....................................... 131 PART 2 Deep network design for medical image reconstruction, synthesis, and selected applications CHAPTER 7 Reconstruction: supervised artifact reduction ..... 137 7.1 Designprinciples .................................. 138 7.1.1 Imagedomainapproaches ...................... 138 7.1.2 Sensordomainapproaches ..................... 139 7.1.3 Dual-domainapproaches ....................... 141 7.2 Casestudy:sparse-viewartifactreduction ............... 142 7.2.1 Background ................................ 143 7.2.2 Methodology ............................... 144 7.2.3 Experiments ................................ 147 7.2.4 Discussion ................................. 149 7.3 Casestudy:metalartifactreduction .................... 150 7.3.1 Background ................................ 152 7.3.2 Methodology ............................... 154 7.3.3 Experiments ................................ 156 7.3.4 Discussion ................................. 162 7.4 Summary ........................................ 162 References....................................... 163 CHAPTER 8 Reconstruction: unsupervised artifact reduction... 169 8.1 Designprinciples .................................. 170 8.1.1 Unpairedlearningapproaches ................... 170 8.1.2 Self-supervisedlearningapproaches .............. 172 8.2 Casestudy:metalartifactreduction .................... 175 8.2.1 Background ................................ 177 8.2.2 Methodology ............................... 177 8.2.3 Experiments ................................ 183 8.2.4 Discussion ................................. 191 8.3 Summary ........................................ 193 References....................................... 194