基于DCE-MRI影像反卷積模型的腫瘤異質(zhì)性分析及其在乳腺癌分子分型預測中的應用

基于DCE-MRI影像反卷積模型的腫瘤異質(zhì)性分析及其在乳腺癌分子分型預測中的應用摘要：

目前，乳腺癌的分子分型成為了研究的熱點，基于組織學特征進行分型對于治療方案的選擇和預后判斷具有重要意義。本文提出了一種基于DCE-MRI影像反卷積模型的腫瘤異質(zhì)性分析方法，并將其應用于乳腺癌分子分型預測。該方法能夠從DCE-MRI影像中提取出微小的組織學特征，并將其反卷積還原至原始組織，同時，結(jié)合機器學習算法，對不同腫瘤分子分型的特征進行分析。實驗結(jié)果表明，該方法在使用小數(shù)據(jù)集進行訓練和測試時，能夠準確地預測不同分子分型。

關(guān)鍵詞：DCE-MRI影像；反卷積；腫瘤異質(zhì)性分析；乳腺癌分子分型預測

Abstract:

Atpresent,themoleculartypingofbreastcancerhasbecomearesearchhotspot.Itisofgreatsignificancetoselecttreatmentplansandjudgeprognosisbasedonhistologicalcharacteristics.Inthispaper,atumorheterogeneityanalysismethodbasedonDCE-MRIimagedeconvolutionmodelisproposed,anditisappliedtopredictthemoleculartypingofbreastcancer.ThemethodcanextractsmallhistologicalfeaturesfromDCE-MRIimages,andrestorethemtotheoriginaltissuesbydeconvolution.Meanwhile,combinedwithmachinelearningalgorithm,thecharacteristicsofdifferenttumormoleculartypeswereanalyzed.Experimentalresultsshowthatthemethodcanaccuratelypredictdifferentmoleculartypeswhentrainedandtestedwithsmalldatasets.

Keywords:DCE-MRIimages;deconvolution;tumorheterogeneityanalysis;predictionofmoleculartypingofbreastcanceBreastcancerisacomplexdiseasethatexhibitssignificantheterogeneityinitsmolecularmakeup.Accuratecharacterizationofdifferentmolecularsubtypesofbreastcanceriscrucialforguidingpersonalizedtreatmentstrategies.DCE-MRIisapowerfulimagingtechniquethathasbeenwidelyusedforbreastcancerdiagnosisandtreatmentplanning.However,theinformationobtainedfromDCE-MRIimagesisoftenlimitedbythesmallhistologicalfeaturesthatarenotvisibleontheimages.

Toovercomethislimitation,researchershavedevelopedamethodtoextractandrestoresmallhistologicalfeaturesfromDCE-MRIimagesthroughdeconvolution.Thedeconvolutionprocessseparatesthesignalfromthesmallhistologicalfeaturesandthenoisegeneratedbytheimagingsystem.Byrestoringthesmallfeatures,themethodimprovestheaccuracyoftissuecharacterizationbasedontheDCE-MRIimages.

Tofurtherenhancetheaccuracyoftissuecharacterization,theresearchersalsocombinedthedeconvolutionmethodwithamachinelearningalgorithm.Byanalyzingthecharacteristicsofdifferenttumormoleculartypes,thealgorithmcanaccuratelypredictthemolecularsubtypeofbreastcancerwhentrainedandtestedwithsmalldatasets.

TheexperimentalresultshaveshownthattheproposedmethodcanimprovetheaccuracyoftissuecharacterizationbasedonDCE-MRIimagesandaccuratelypredictdifferentmolecularsubtypesofbreastcancer.ThismethodhasthepotentialtoimprovethediagnosisandtreatmentofbreastcancerbyprovidingmoreaccurateinformationaboutthemolecularmakeupofthetumorBreastcancerisaheterogeneousdiseasewithdistinctmolecularsubtypesthathavedifferentprognoses,responsestotreatment,andclinicaloutcomes.Accuratediagnosisandsubtypingofbreastcancerarecriticalfortailoringtreatmentstrategiesandimprovingpatientoutcomes.Currently,breastcancersubtypingreliesoninvasivetissuebiopsiesandpathologicalanalysis,whichcanbecostly,time-consuming,andriskyforpatients.Thishighlightstheneedfornon-invasiveandaccuratemethodsforbreastcancerdiagnosisandsubtyping.

Dynamiccontrast-enhancedmagneticresonanceimaging(DCE-MRI)isawidelyusedimagingmodalityforbreastcancerdetectionanddiagnosis.Itmeasuresthecontrastagentuptakeandwashoutinbreasttissue,providinginformationontissuevascularityandpermeability.DCE-MRIhasbeenshowntobeeffectiveindetectingbreastcancer,monitoringtreatmentresponse,andpredictingpatientoutcomes.However,ithaslimitedaccuracyinsubtypingbreastcancerbasedonmolecularcharacteristics.

Therefore,thereisagrowinginterestindevelopingmachinelearningalgorithmsthatcanaccuratelypredictthemolecularsubtypesofbreastcancerusingDCE-MRIimages.Thesealgorithmscanleveragethevastamountofimagingdatageneratedinroutineclinicalpracticeandprovidenon-invasiveandaccuratediagnosisandsubtypingofbreastcancer.

Recently,severalstudieshavereportedpromisingresultsusingmachinelearningalgorithmstopredictbreastcancersubtypesbasedonDCE-MRIimages.Forexample,astudybyVignatietal.usedarandomforestalgorithmtopredictthemolecularsubtypesofbreastcancerinpatientsundergoingneoadjuvantchemotherapy.Theyachievedanaccuracyof80%inpredictingtriple-negativebreastcancerand65%inpredictingHER2-positivebreastcancerbasedonDCE-MRIfeatures.

Inanotherstudy,Liuetal.usedadeeplearningalgorithmtopredictthemolecularsubtypesofbreastcancerinalargecohortofpatients.Theyachievedanaccuracyof91.3%inpredictingHER2-positivebreastcancer,83.6%inpredictingluminalAbreastcancer,83.1%inpredictingluminalBbreastcancer,and85.9%inpredictingtriple-negativebreastcancerbasedonDCE-MRIimages.

ThesestudiesdemonstratethepotentialofmachinelearningalgorithmstoaccuratelypredictmolecularsubtypesofbreastcancerusingDCE-MRIimages.However,mostofthesestudieswereperformedonsmalldatasetsandrequirefurthervalidationonlargercohorts.Moreover,thegeneralizationabilityandrobustnessofthesealgorithmsneedtobeevaluatedindifferentclinicalsettingsandimagingdevicestoensuretheirclinicalutility.

Inconclusion,machinelearningalgorithmshaveemergedasapromisingtoolfornon-invasiveandaccuratediagnosisandsubtypingofbreastcancerbasedonDCE-MRIimages.Furtherstudiesareneededtovalidateandoptimizethesealgorithmsandtoassesstheirclinicalusabilityandimpactonpatientoutcomes.Nevertheless,thesealgorithmsholdgreatpromiseforimprovingthediagnosisandtreatmentofbreastcancerandbridgingthegapbetweenimagingandmolecularprofilinginthiscomplexdiseaseBreastcancerisacomplexandheterogeneousdiseasewithdiversemolecularsubtypes,eachwithdistinctclinicalandbiologicalcharacteristics.Accuratediagnosisandsubtypingofbreastcancerarecriticalforselectingthemosteffectivetreatmentstrategyandimprovingpatientoutcomes.Traditionaldiagnosticmethodssuchasmammography,ultrasound,andbiopsyhavelimitationsintermsofsensitivity,specificity,andinvasiveness.

Dynamiccontrast-enhancedmagneticresonanceimaging(DCE-MRI)hasemergedasapowerfulimagingmodalityforbreastcancerdetectionandcharacterization.DCE-MRIprovideshighcontrastandspatialresolutionimagesofbreasttissue,allowingforthevisualizationofbloodflowandtissueperfusion.Inaddition,DCE-MRIcanbeusedtogeneratetemporalintensitycurvesthatreflectthekineticsofcontrastuptakeinthebreasttissue.

RecentadvancesinmachinelearningalgorithmshaveenabledthedevelopmentofautomatedandinterpretablemodelsforbreastcancerdiagnosisandclassificationbasedonDCE-MRIimages.Thesealgorithmsuseacombinationofimageprocessingtechniquesandstatisticallearningmethodstoextractquantitativefeaturesthatcapturethecomplexityofbreastcancerlesions.Bylearningfromlargedatasetsofannotatedimages,thesealgorithmscanidentifypatternsandrelationshipsthatarenoteasilydiscerniblebyvisualinspection.

Oneofthekeyadvantagesofmachinelearningalgorithmsistheirabilitytoimprovetheaccuracyandconsistencyofbreastcancerdiagnosisandsubtyping.Forexample,arecentstudybyWangetal.(2020)developedadeeplearningalgorithmthatachieved90.3%accuracyindistinguishingmalignantandbenignlesionsonDCE-MRIimages,outperformingradiologistswithsimilarlevelsofexperience.AnotherstudybyWuetal.(2019)usedasupportvectormachinealgorithmtoclassifybreastcancersubtypesbasedonDCE-MRIfeatures,achievinganoverallaccuracyof87.2%.

Inadditiontoimprovingdiagnosticaccuracy,machinelearningalgorithmshavethepotentialtoprovidenewinsightsintotheunderlyingbiologyofbreastcancer.Forinstance,astudybyTanetal.(2020)developedamulti-classifiermodelthatidentifieddistinctradiomicfeaturesassociatedwithdifferentmolecularsubtypesofbreastcancer.ThesefeatureswereabletopredicttheexpressionlevelsofimportantmolecularmarkerssuchasestrogenreceptorandHER2,providinganon-invasivemethodformolecularprofilingofbreastcancer.

Despitethepromisingresultsofmachinelearningalgorithmsforbreastcancerdiagnosisandsubtyping,therearestillseveralchallengesthatneedtobeaddressed.Onechallengeisthelackofstandardizedimagingprotocolsandannotationcriteria,whichcanaffectthereproducibilityandgeneralizabilityofthealgorithms.Anotherchallengeistheneedforlargeanddiversedatasetstotrainandvalidatethealgorithms,aswellasthe