版权说明:本文档由用户提供并上传,收益归属内容提供方,若内容存在侵权,请进行举报或认领
文档简介
NatureBiomedicalEngineering
naturebiomedicalengineering
Article
/10.1038/s41551-026-01637-3
AgeneralizabledeeplearningsystemforcardiacMRI
Received:16October2024
Accepted:23February2026
Check
forupdates
RohanShad1,CyrilZakka2,DhamanpreetKaur2,MrudangMathur2,
RobynFong2,JosephCho2,RossWarrenFilice3,JohnMongan4,
KimberlyKallianos4,NishithKhandwala5,DavidEng5,MatthewLeipzig2,
WalterRWitschey6,AlejandrodeFeria7,VictorAFerrari7,EuanAAshley8
...,MichaelA.Acker1,CurtisLanglotz9&WilliamHiesinger2
CardiacMRIallowsforacomprehensiveassessmentofmyocardial
structure,functionandtissuecharacteristics.Herewedescribea
foundationalvisionsystemforcardiacMRI,capableofrepresentingthe
breadthofhumancardiovasculardiseaseandhealth.Ourdeep-learning
modelistrainedviaself-supervisedcontrastivelearning,inwhichvisual
conceptsincine-sequencecardiacMRIscansarelearnedfromtheraw
textoftheaccompanyingradiologyreports.Wetrainandevaluateour
modelondatafromfourlargeacademicclinicalinstitutionsintheUnited
States.Weadditionallyshowcasetheperformanceofourmodelson
theUKBioBankandtwoadditionalpubliclyavailableexternaldatasets.
Weexploreemergentcapabilitiesofoursystemanddemonstrate
remarkableperformanceacrossarangeoftasks,includingtheproblem
ofleft-ventricularejectionfractionregressionandthediagnosisof39
differentconditionssuchascardiacamyloidosisandhypertrophiccar-
diomyopathy.Weshowthatourdeep-learningsystemiscapableofnot
onlycontextualizingthestaggeringcomplexityofhumancardiovascular
diseasebutcanbedirectedtowardsclinicalproblemsofinterest,yielding
impressive,clinical-gradediagnosticaccuracywithafractionofthetrainingdatatypicallyrequiredforsuchtasks.
Offeringunparalleleddiagnosticclarity,cardiacmagneticresonanceimaging(CMRorcardiacMRI)isthereferencestandardforassessingcardiacanatomyandfunction
1
,
2
.Dependingonthetechniqueused,itenablesclinicianstocapturevideographicsequencesofcardiacandvalvularmotion,quantifyscarringortissueoedemawithinthecardiacmusculature,andidentifyregionsofpoorperfusion—allwithoutexposuretoionizingradiation
2
.Despitethiswealthofdataavailable,deep-learningsystemscapableoflearninghigh-quality
representationsofhumancardiacdiseasefromCMRhavenotyetbeendescribed.
Deeplearninghasshownincrediblepromiseinthediagnosisofcardiovasculardiseasewithelectrocardiography(EKG),retinalscansandechocardiography
3
–
7
.Traditionally,thesenetworksaretrainedtodetectahandfulofpre-definedandcurated‘disease’conditionsinthebackgroundofnegative‘normal’cases.Whilesuperficiallysuc-cessful,manyofthesesystemsfailwhentestedonreal-worldclinical
1DivisionofCardiovascularSurgery,DepartmentofSurgery,UniversityofPennsylvania,Philadelphia,PA,USA.2DepartmentofCardiothoracicSurgery,StanfordUniversity,Stanford,CA,USA.3DepartmentofRadiology,MedstarGeorgetownUniversityHospital,Washington,DC,USA.4Department
ofRadiologyandBiomedicalImaging,UniversityofCalifornia,SanFrancisco,CA,USA.5BunkerhillHealth,SanFrancisco,CA,USA.6Department
ofRadiology,UniversityofPennsylvania,Philadelphia,PA,USA.7DivisionofCardiovascularMedicine,DepartmentofMedicine,Universityof
Pennsylvania,Philadelphia,PA,USA.8DivisionofCardiovascularMedicine,DepartmentofMedicine,Genetics,andBiomedicalDataScience,
StanfordUniversity,Stanford,CA,USA.9DepartmentofRadiology,Medicine,andBiomedicalDataScience,StanfordUniversity,Stanford,CA,USA.
e-mail:
rohan.shad@
;
willhies@
Article
/10.1038/s41551-026-01637-3
NatureBiomedicalEngineering
datathatareoftenheterogeneouswithnumerousconcomitantabnor-malfindings
8
.Patientswithinheritedcardiomyopathies,forexam-ple,maypresentwithseverevalvulardisease;thosewithevidenceofventricularthrombus,ontheotherhand,mayhavesevereheartfailurefromremoteischaemicinsults.Withthetraditionalsuperviseddeep-learningapproach,itremainschallengingtocontextualizethisdiversityofdiseasepresentationfromCMRscans,withparameterslearnedforoneproblemrarelygeneralizingtoothers
9
.Thesesystems,therefore,mustbere-trainedfromscratchforeverynewclinicaltaskofinterest,requiringthousandsoftrainingexampleseachtime.Unlikehumanclinicians,deep-learningmodelsdonothaveabaselinefoun-dationofclinicalandpathophysiologicalunderstandingoverwhichlearningspecifictaskscanbeaccelerated.Ultimately,thishasrestrictedresearchinthefieldtotasksthateitherautomateobjectivestructuralmeasurementsortodiseasesthataresimplymoreprevalent
10
.
Herewedescribeatransformer-basedvisionsystemthatlearnscomplexpathophysiologicalvisualrepresentationsfromalargemulti-institutionaldatasetof19,041CMRscans,guidedbynaturallanguagesupervisionfromthetextreportsaccompanyingeachCMRstudy.Weusealargelanguagemodeltohelp‘teach’avisionnetworktogeneratemeaningfullow-dimensionalrepresentationsofCMRstudies,byshowingexamplesofhowradiologistsdescribewhattheyseewhiledraftingtheirreports.Wedescribehowthesemodelsgrouptogetherpatientswithsimilarpathophysiologicalandevensocio-demographiccharacteristics,withnoexplicitsupervisionforthesetasks.Wefinetunethisdeep-learningsystemonsmallerexpert-labelleddatasetsforarangeofclinicallyrelevanttasksrangingfromleft-ventricularejectionfractionestimationtothedetectionof39differentcardiovas-cularconditions.Wevalidateoursystemonmultipleexternaldatasetsfromdistinctgeographicalsitesandhealthsystems.Theresultisageneralizableanddata-efficientCMRdeep-learningsystemcapableofrepresentingthebreadthofcardiovasculardiseaseandhealth.Figure
1
detailsanoverviewoftheproject.
CMRimagingdata
Theinputstoournetworksaresteady-statefreeprecession(SSFP)cinesequencestakenalongmultipledifferentcardiacviewplanes.Typically,SSFPsequencesarecharacterizedbyahighbloodsignalintensity,withlowerintensitymyocardialsignal
11
,
12
.Thistechniqueallowsfortheacquisitionofhigh-contrastdynamic-motionscansoftheheartduringabreathholdtimedwithEKGgatingtocaptureimagesbetweenheartbeats.Thesecine-sequencevideosarecapturedalongvariousspatialviewplanesasdescribedbelow.Ourdataaresourcedfromaheterogeneoussamplepopulationwithsubstantialvariabilityinacquisitionandscannervendors,withtrainingdatasourcedfromStanfordMedicine,MedStarandUCSF.AdditionaldatafromtheUKBioBank,KaggleGrandDataChallenge,AutomatedCardiacDiagnosisChallenge(ACDC)datasetandtheUniversityofPennsylvaniaweresourcedtorigorouslyassessmodelgeneralizability.Specificdescrip-tionsofeachcohortareprovidedinMethodsandSupplementaryFig.1.
Differentcardiacstructuresmaybevisualizedbetterinoneviewplanethaninanother.Asaresult,findingsindicativeofpathologymayonlybevisibleinaspecificMRIviewplane.Thisisconceptuallyanalogoustothechallengesfacedindeeplearningforhistopathology,whereonlyafractionofthebiopsiedtissuemightcontainfindingsdiagnosticofdisease
13
.Welimitourselvestoshort-axisstacks(SAX),four-chamber(4CH),three-chamber(3CH)andtwo-chamber(2CH)viewsforthisworkasthesearemostconsistentlycapturedirrespectiveofquerypathology.DemographicdataavailableforcertainsubsetsofthedataaredetailedinSupplementaryTable1.
Pre-trainingframeworkandevaluationoflow-dimensionalrepresentations
Thetraditionalapproachfordeeplearningwithcardiovascularimaginghasbeentoassignsomelabeltoeachimagingscanviatediousmanual
annotation.Large-scalelabellingoftrainingdatarequiresextensivedomainknowledgeandclinicalexpertise,butultimatelyconstrainsthemodelstofeaturesexplicitlylabelledinthedataset.Modelstrainedthiswayrequirethousandsofexamplesforadequateperformanceandareunabletoaccountforfindingsoutsidethescopeofwhattheyweretrainedtoidentify.Recentadvancesinself-superviseddeeplearninghaveshownpromiseinreducingthisrelianceonvastquantitiesofexpert-labelleddataacrossawiderangeofmodalitiesrangingfrompathologyslidestochestX-rays
14
–
21
.Aframeworkwhereinunstruc-turedtextcouldbeusedasamethodofself-supervisionforachestX-raydeep-learningsystemhasbeendescribed
16
.Inthismethodofcontrastivelearning,twoneuralnetworksareusedtoproduceapairoflow-dimensionalrepresentationsforeachpairofcontextuallyrelatedinputsfromtwoseparatemodalities
22
.WeextendtheseconceptstothespatiotemporalandmultiviewproblemofcardiacMRI.Duringthepre-trainingprocess,thenetworksaretrainedtomatchtruepairsoftextreportandMRIscansbyoptimizingforacontrastiveobjec-tive
22
.Sincethereleaseofourinitialpreprint,othershavereportedpromisingresultsusingsimilarcontrastive-learningmethodstotraingeneralizable‘foundationmodels’forcomputedtomography(CT)andechocardiography
23
–
25
.
Visualfeaturesofaparticulardiseaseareiterativelyassociatedwithtextualcues,withatrainingprocessguidedbyarichunstruc-tureddescriptionofdiseasecreatedbycliniciansaspartofroutineclinicalworkflow.Weuseanimplementationofamultiscalevisiontransformer(mViT)fortheCMRvisionencoderandabidirectionalencoderrepresentationsfromtransformers(BERT)texttransformerfortheMRIreportencoder
26
,
27
.Wepre-trainandvalidateon14,073cardiacMRIscans(12,707uniquepatients)fromStanford,UCSFandMedStar.SpecificimplementationdetailsaredescribedinMethods.
Weplottheembeddingsgeneratedbythevisionencoderonthevalidationsetusingstandarddimensionalreductionalgorithms(Meth-ods),demonstratingtheprogressiveemergenceoflocalandglobalstructureinthetwo-dimensional(2D)projectionsoftheseembeddingsastrainingprogresses(Fig.
1
).Oncetrainingiscomplete,thenetworkisfrozenanddeployedontoapubliclyavailableexternaldatasetofcine-CMRsequencessourcedfromFrance(ACDCdataset)
28
.WeplottheembeddingsforeachstudyintheACDCdatasetontotwodimen-sionsandconfirmthatcomparedtoabaselinenetworkofthesamearchitecture,ourvisionsystemisabletoseparatedifferentdiseaseconditionswithremarkableconsistency(SupplementaryFig.4).Thisisdespitethelackofanydirectedsupervisionduringthetrainingprocessinseparatingconditionssuchashypertrophiccardiomyopathyfromdilatedcardiomyopathyorrightventriculardysfunction.
ValidationontheUKBioBank
Wesourced159,883cine-CMRscansrepresenting45,623participantsfromtheUKBioBankandassessedtheperformanceoflarge-scalecontrastivepre-trainingonadatasetrepresentingarelativelyhealthyparticipantpopulationfromageographicallydistinctcontinent.Wehypothesizedthatifusefulrepresentationsarelearnedbyourdeep-learningsystemduringthepre-trainingphase,itshouldbetrivialtoseparateparticipantsonthebasisoffeaturesindicativeofdiseasewithoutanyadditionalsupervisedtraining.Weuseourcontrastivepre-trainednetworkandfreezetheweightstopreventanyadditionallearning,followingwhichwepasseachavailablecine-CMRsequencethroughthenetwork.Thenetworkgeneratesalow-dimensionalembeddingofeachinput.Weprocesseachviewseparatelyandstoretheembeddingsgeneratedforfurtherprocessing.AswiththeexperimentsontheACDCdataset,werepeattheexperimentwithanidenticalnetworkwithweightsinitializedfromatrainingrunontheKinetics-400actionrecognitiondataset,avideodatasetofnaturalsceneryandactivities
29
.
Weelectedtousetheunsupervisedt-distributedstochasticneighbourembedding(t-SNE)algorithmtodimensionallyreducethe
Article
/10.1038/s41551-026-01637-3
aLarge-scalecontrastivepre-training
Videotransformer(36milliontrainableparams)
Theleftventricleexhibits
di仟usedilatationand
hypokinesia,most
pronouncedattheapex.Despiteapicalinferior
hypokinesia,thereisno
evidenceofthrombus
formation.Theright
ventricleisnormalin
size,wallthickness
Texttransformer
b
15
UMAP−2
UMAP−2
10
5
0
20
10
V
T
1
T
2
T
3
T
n
V1V2V3Vn
VT
11
VT
21
VT
31
VT
n1
VT
12
VT
22
VT
32
VT
13
VT
23
VT
33
VT
1n
VT
nn
T
0
Epochnumber:1
UMAP−2
−5051015
UMAP−1
Epochnumber:20
UMAP−2
−5051015
Epochnumber:6
Epochnumber:2
15
10
15
UMAP−2
5
10
0
5
−5
0
051015
UMAP−1
Epochnumber:50
−50510
UMAP−1
Epochnumber:100
15
10
15
UMAP−2
5
10
0
5
−5
0
−50510
−5051015
−5
UMAP−1
UMAP−1
UMAP−1cZero-shotevaluation
0trainableparams
t-SNE−2
0
Patient001
–3
–2.50
3
DiseaselabelNormal
MI
RVdysfunctionHCM
DCM
–6
2.5
t-SNE−1
dFinetuning
600ktrainableparams
Patient001
Ischaemic
cardiomyopathy
LVEF%regressionamyloidosis
Fig.1|Projectoverview.a,Large-scalecontrastivepre-training.CMRcine
sequencesintheformofvideodataarefedtoavideotransformernetwork,
andthepairedtextreportsarefedintoaparalleltexttransformernetwork.
Theembeddingsproducedbyeach(V,video;T,text)representcompressed
representationsofthecomplexinputstoeachnetwork.ThenetworksaretrainedtomaximizeagreementbetweentruepairsofvideoandtextthatarisefromthesameCMRscan(thatis,thediagonalofthesimilaritymatrixhighlighted:V1T1,
V2T2…VnTn)andminimizetheagreementbetweensamplesthatcomefromdifferentscans.b,Intheprocessina,weseethatastrainingprogresses,morecomplexlocalandglobalstructuralfeaturesemergeinthevideoembeddings(V).Thescansinitiallyviewedasanindistinctamalgaminhigh-dimensionalspaceatthestartoftraining(epoch#1)begintodevelopseparationsand
localizedclustersthataremoredistinctbyepoch#100.c,Evaluatingtheabilityofthenetwork,onexternaldatasets,inseparatingdiseasesorprobingtheUK
BioBankasdescribedinlatersectionsisachievedbyfreezingalltheparametersofthevideoencoder(atthispoint,thetextnetworkisdiscarded).Embeddingsproducedfromdifferentviewscanthenbeplottedin2Dviastandarddimensionreductionalgorithms.d,Finally,leavingthelastlayerunfrozen(~600,000
trainableparameters),weshowcasedata-efficientfinetuningtowardsspecificclinicaltasksofinterest.Ofnote,theembeddingsarefedintoasecondary
networkdesignedtoaggregateinformationfromdifferentviewsintoasingleprediction,withafewexampleslisted:regressionofleft-ventricularejectionfraction(LVEF%)anddiagnosisofamyloidosisorischaemiccardiomyopathy.Asinglevisionmodelcanthusbeappliedtoawiderangeofdifferenttasks.
embeddingsgeneratedforeachviewandvisualizethemintwodimen-sions
30
.Weexpectedthenetworksatbaselinetoatleastbecapableofgeneratingembeddingsofsufficientqualitytoseparatedifferentviewswhenvisualizedwitht-SNE,astheyareobviouslydistincteventotheuntrainedeye.InFig.
2
(left),weshowtheunsupervisedclusteringofembeddingsproducedbyourcontrastivepre-trainednetworkin2Dt-SNEspace,forall45,623UKBioBankparticipants.Wefindthatthepre-trainednetworkisessentiallyviewinvariant.Keytodrivinginvari-ancebasedonviewplaneswastoensurethatallviewplanesfromthesamestudyarealignedwithatextembeddingfromthesametextreportduringpre-training.Thisallowsforattentiontobeplacedinsteadon
featuresofimportancehighlightedinthetext.Weseearelativelydenseclusterofpatientswithejectionfractions<35%.Surprisingly,withoutanyexplicitinstruction,wefindthatcontrastivepre-trainingallowedforsharpdelineationofsexandseparationbyage.Itislikelythatthepresenceoftextualinformationcontainingphrasessuchas‘adultfemale’withinsomeoftheMRIreportsallowedthevisionnetworkstolearnfeaturesthatidentifydemographiccharacteristics.Importantly,allofthesefindingsareseenwiththenetworksentirelyfrozen,withnoadditionallearningormodelparameterupdatespossible.ComparisonwithsimilarexperimentsusingtheKinetics-400pre-trainedbaselineisinstructiveindemonstratingthelackofanysuchbiomedicallatent
Article
/10.1038/s41551-026-01637-3
Kineticsbaseline
Contrastivepre-trained
t-SNE_2
t-SNE_2
25
Sex
Female
Male
0
_25
_40_2002040
t-SNE_1
25
0
_25
_50
_25025
t-SNE_1
Sex
t-SNE_2
t-SNE_2
Yearofbirth
1970
1960
1950
1940
_40_2002040
t-SNE_1
25
0
_25
Ejtifti
25
0
_25
_50
_25025
t-SNE_1
Age
t-SNE_2
t-SNE_2
_40_2002040
t-SNE_1
>35%
<35%
75
50
25
25
0
_25
25
0
_25
_50
_25025
t-SNE_1
econracon
t-SNE_2
t-SNE_2
_25
_40_2002040
t-SNE_1
25
0
No
Yes
0
_25
_50
_250
Cardiomyopathystatus
t-SNE_1
25
25
t-SNE_2
t-SNE_2
25
0
_25
_40_2002040
t-SNE_1
Views
.2CH
.3CH
拳4CH
SAX
25
0
_25
_50
_25025
t-SNE_1
Views
Fig.2|Contrastivepre-trainingenableslearningofdemographicand
pathophysiologicalrepresentations.2Dt-SNEplotsoflow-dimensional
embeddingsgeneratedfromaforwardpasson34,490uniquescansfromtheUKBioBank(n=31,693uniqueindividuals)usingaKinetics-400pre-trainedcheckpoint(baseline,rightcolumn)comparedtoacontrastivepre-trained
checkpoint(ours,leftcolumn).Thecolourlabelshelpwithunderstanding
thebasisofclustering.WhileatfirstglanceitmayappearthattheKinetics
baselinemodelproducesasetof5readilyseparableclustersvsthemore
homogeneous-appearingplotforthecontrastivepre-trainedmodel,the
basisofthebaselineKinetics-initializedclustersaresimplythevariousMRI
viewplanes.OnfurtherexplorationwefindthattheKinetics-400-generatedembeddingsfailtocapturetherequiredinformationtoseparatelow-ejectionfractionstates,cardiomyopathy,genderorage(lackofseparationofcolourforeachvariable).Thisisdifferentfromthecontrastivepre-trainedgeneratedembeddingsthatshowcleardemarcationsbyphysiologicallyrelevant
characteristics,allowingforzero-shotseparationoflow-ejectionfraction
states,cardiomyopathy,genderandage.Colourlegendsforeachsubplotareshownontherightofeacht-SNEfigure.Viewinvarianceisabuilt-infeatureofthecontrastivepre-trainingprocess,evidencedbythecharacteristicabsenceofclusteringbyviewplane.
knowledgeviatraditionaltrainingmethods(Fig.
2
right).NodistinctclustersbeyondthoseofthemajorMRIviewplanesstudiedareseenwithoutourcontrastivepre-trainingmethods.
Self-attention-basedaggregationofembeddingsfrommultipleCMRviews
Weperformaseriesofexperimentstodefinetheimpactofcontrastivepre-trainingwhenthesemodelsaretaskedtoavarietyofclinicallyrel-evantregressionandclassificationproblems.EachviewinacardiacMRIstudymaybeofvaryingdiagnosticutility,dependingonthepresent-ingpathology.TomimictheapproachofcliniciansreportingcardiacMRIscans,asecondaryneuralnetworkisusedtoaggregateusefulinformationfromeachviewplanetogenerateafinalscanleveloutput(Methods).Asaresult,ourmodelsuseallavailableviews(4CH,2CH,3CHandSAX)whileanalysingeachpatientexamination.WetreatthepresenceofmultipleMRIviewplanesasaweaklylabelledmulti-instancelearningproblem.Videosfromeachviewplaneareprocessedviaourpre-trainedvideotransformerencoder,andtheresultingembeddingsarefedintoaself-attention-basedmulti-instancelearningmodule.Theself-attentionmoduleistrainedtoidentifyrelevantfeaturesacrosseachavailableviewplane,independentoftheactualnumberofviewssup-plied
31
.Therelativecontributionofoneviewoveranothertowardsthefinaloutputiscalculatedviaalearnedweightedaverageoftheoutputembeddings.Thisassignsimportancetodifferentviewsdependingonthepathologicalfeaturespresentinthem.Thisisconceptuallyidenticaltoassigninghigherimportancetoaselectionofzoomed-inpatchesofhistopathologywhole-slideimagesthatcontainfeaturesofmalignancy,inthebackgroundofalargenumberofnon-diagnosticpatches
13
,
31
.
Automatedestimationofleft-ventricularejectionfraction
Wedefinetheimpactofcontrastivepre-trainingontheproblemofpredictingleft-ventricularejectionfraction(LVEF),acommonlycalculatedmetricofcardiacfunction,ontwoexternaldatasets:theUKBioBankreferencedaboveandapubliclyavailableCMRdataset(Kaggle)forthisproblem.Deep-learning-derivedautomatedLVEFmeasurementsinechocardiographyandcardiacMRIaretypicallybasedonsegmentationmodelsorhybridregression-segmentationsystems
4
,
28
,
32
,
33
.Suchnetworksaretrainedtooutlineleftventricularchambersduringendsystoleandenddiastole,andassuchhavenofurtherinherentunderstandingofthediseasespace.Nonetheless,thesenetworksperformexceptionallywellinthetaskofLVEFestimationastheyreplicatethemeasurementworkflowofroutineclinicalpractice.Forinstance,ameanabsoluteerror(MAE)of3.2hasbeenreportedforthetaskofpredictingLVEFinasubsetoftheUKBioBankparticipantsusingshort-axissequences
34
.Tworecentpapersreleasedwhileourmanuscriptwasinreviewarenotablefortheirtechnicaladvanceintheareaofdeep-learning-poweredLVEFestimation:ref.
35
describesCineMA,amasked-autoencodersystemutilizingahybrid-segmentationsupervision;andref.
36
describesmaskedautoencoderscombinedwithacontrastiveobjective.WhilebothachievecompetitiveperformanceonLVEFestimationintheUKBioBank(CineMAwithaMAEof3.34%andref.
36
withaMAEof2.95%),critically,thesesystemswerebothpre-trainedandfinetunedontheUKBioBankitself,makingdirectcomparisonsofgeneralizabilitychallenging.
Prospectivestudieshaveshownthatthereisconsiderablevari-abilityininstitutionalanddataset-specificprotocolsforcalculationofcertainmetrics,butclinicianscantypicallybeexpectedtomakeestimatesofLVEFwithinBland–Altmanlimitsof_12%to+12%
37
,
38
.Withourpre-trainedvisionencoderfrozen,barringthelastlinearlayer,wefinetunethenetworkon34,488uniqueCMRscansfromtheUKBioBankforthetaskofpredictingLVEF.Thevisionencodercannolongerlearndataset-specificfeatureswiththisapproachandmustrelyonlearnedrepresentationalabilitiesfrompreviouspre-training.Toincorporateinformationfromeachavailableviewtoproducea
Article
/10.1038/s41551-026-01637-3
NatureBiomedicalEngineering
‘patientlevel’predictionofLVEF,weuseamulti-instanceself-attentionregressionheadasdescribedaboveandinputcine-CMRsequencesfromallavailableviewswithoutanyadditionalqualitycontrolsteps(Methods).Onahold-outtestsubsetofUKBioBankparticipantsusingourapproach,wereportaMAEof3.344(s.d.3.615),withBland–Altmanlimitsofagreementof_9.91%to+9.61%.Atbaseline,usingatradition-allytraineddeep-learningsystemofthesamemodelarchitectureasourcontrastivepre-trainedmodels,wereportanMAEof4.603(s.d.4.409),withBland–Altmanlimitsofagreementof_12.15%to+12.8%.
Thesemetricsrivalthoseofhand-crafteddeep-learningsystemstocalculateLVEFusingsegmentationmasksonmanuallyselectedend-systolicandend-diastolicframes,andarewellwithinerrorlimitsofcliniciansfollowingstrictannotationprotocols
35
,
37
.Wefreezethisfine-tunedmodelandevaluateitoncardiacMRIscansfromapub-liclyavailabledataset,wherepatientswererecruitedfromhospitalsystemsbasedintheUnitedStates(KaggleDataScienceBowl)
39
.TheKaggleexternaldatasetcontainsalargerproportionofpatientswithdiseasedheartsandadditionallyutilizesaslightlydifferentscanningprotocolandmethodfordeterminationofground-truthLVEFlabels.OnthisdatasetwereportaMAEof6.880(s.d.5.309),withBland–Alt-manlimitsofagreementfo_18.7%to+8.03%.GiventhedifferencesinannotationmethodscomparedtodatafromtheUKBioBank,thefrozenfine-tunednetworkshowedamodestsystematicunderpredictionofLVEFby5.36%(95%CI4.86–5.87)(Fig.
3b
).Diagnosticplotsfortest-setresultsandadditionalresultswithimprovementsontheKaggledatasetwithbiascorrectionareavailableinSupplementaryFigs.7and8(finalbias-correctedMAE4.861),showingsuperiorperformanceinthehighandlowLVEFrange,withsmallerresidualerrorswhenmodelsareinitial-izedfromcontrastivepre-trainedweights.Finally,wemanuallyreviewCMRstudiesformodelpredictionswiththelargestabsoluteerrorsandfindthatthemajorityofthesestemfromincorrectground-truthlabelsorartefact/noise-degradedimages(SupplementaryFig.9andTable4).
Nevertheless,ifthesepredictedLVEFvaluesweretobeusedforidentifyingpatientswithheartfailurewithreducedejectionfraction(HFrEF)<40
温馨提示
- 1. 本站所有资源如无特殊说明,都需要本地电脑安装OFFICE2007和PDF阅读器。图纸软件为CAD,CAXA,PROE,UG,SolidWorks等.压缩文件请下载最新的WinRAR软件解压。
- 2. 本站的文档不包含任何第三方提供的附件图纸等,如果需要附件,请联系上传者。文件的所有权益归上传用户所有。
- 3. 本站RAR压缩包中若带图纸,网页内容里面会有图纸预览,若没有图纸预览就没有图纸。
- 4. 未经权益所有人同意不得将文件中的内容挪作商业或盈利用途。
- 5. 人人文库网仅提供信息存储空间,仅对用户上传内容的表现方式做保护处理,对用户上传分享的文档内容本身不做任何修改或编辑,并不能对任何下载内容负责。
- 6. 下载文件中如有侵权或不适当内容,请与我们联系,我们立即纠正。
- 7. 本站不保证下载资源的准确性、安全性和完整性, 同时也不承担用户因使用这些下载资源对自己和他人造成任何形式的伤害或损失。
最新文档
- 2026年办公室文员试题及答案
- 家长委员会发言稿(资料15篇)
- 工作失误检讨书资料
- 2026年湖南益阳市中考政治考试真题及答案
- 2026年保密教育测试题库试题附答案完整版
- 高中语文人教统编版选择性必修 中册4.1 修辞立其诚教案
- 第十五课 在挫折中成长教学设计小学心理健康南大版六年级-南大版
- 初中地理第二节 世界的语言和宗教教学设计
- 船舶服务协议书范本
- 第3节 单摆教学设计高中物理鲁科版选修3-4-鲁科版2004
- 净菜加工的行业分析报告
- 公文写作业务培训课件
- 牧运通官方兽医试题题库带答案详解(满分必刷)
- 2025年专升本考试真题及答案语文
- 2025年绿色信贷流程
- 业务连续性培训课件
- 肺癌影像学诊断规范
- 升压站砌筑工程施工方案
- 通信工程项目验收与质量管理考试题
- 智能楼宇管理员培训试题及答案
- 中航工业中层竞聘笔试必刷题
评论
0/150
提交评论