您好,欢迎访问籍索

上传文档

当前位置:首页 > 测试分类 > B751_9145

B751_9145

  • 管理员
  • 7 次阅读
  • 1 次下载
  • 2022-09-28 17:30:15
二扫码支付 微信
二扫码支付 支付宝

还剩... 页未读,继续阅读

免费阅读已结束,点击付费阅读剩下 ...

¥ 10 元,已有7人购买

付费阅读

阅读已结束,您可以下载文档离线阅读

¥ 10 元,已有1人下载

付费下载
文档简介:

AMulti-scaleDeepBottleneckNetworkforECGSmallSampleClassification1stDaweiLiSouth-CentralUniversityforNationalitiesCollegeofelectronicsandInformationEngineeringWuhan,Chinaleedavidhust@outlook.com3rdCongLiuSouth-CentralUniversityforNationalitiesCollegeofElectronicsandInformationEngineeringWuhan,Chinaliucong16948@foxmail.com2ndLiqiLiaoSouth-CentralUniversityforNationalitiesCollegeofelectronicsandInformationEngineeringWuhan,Chinaliaoliqi2022@163.com4thXiaoweiXuGuangdongAcademyofMedicalSciencesGuangdongProvincialPeople’sHospitalGuangzhou,Chinaxiao.wei.xu@foxmail.comAbstract—Electrocardiogram(ECG)playsacrucialroleinthediagnosisofcardiovasculardiseases.Inrecentyears,theextensiveuseofdeeplearningalgorithmprovidesanopportunitytoimprovetheefficiencyofdiagnosis.However,theexistingstudieshavemainlyfocusedontheECGclassificationusingsingle-scaleinformation.Formostsmallsamplesapplications,thelimitedtrainingdatacannotprovideseveralsingle-scaleinformationtodemonstratethedistributionofdiseaseclasses.Inthispaper,weproposeamulti-scaledeepbottlenecknetwork,whichcanbeusedtodetectcardiovasculardiseasesespeciallyarrhythmiasinsmallECGdataset.Inthismodel,wedesignedthreebranchescombiningthebottleneckstructurestoextractECGfeaturesatdifferentscales.EachbottleneckstructureisaddedwithaSqueezeandExcitationNetwork(SE-Net)toenhancetheeffectivefeaturechannelsandsuppresstheinefficientones.Toverifytheperformanceofthemodel,wecomparedthe12-leadsECGdatasetcollectedfromGuangdongProvincialPeople’sHospitalwiththatofthepublicMIT-BIHarrhythmiadataset.Forthesetwodatasets,98.2%and98.7%accuracyareachieved,respectively,whichcanexceedtheperformanceofcurrentexistingmethods.ExperimentalresultsshowthatthisworkisespeciallybeneficialtoimprovetheclassificationaccuracyofsmallECGdatasetsandcanbeevenappliedtoothertasksofsmalldatasetsuchasobjectdetection.IndexTerms—Smallsampleclassification,Cardiovascular,Multi-scale,BottleneckStructureI.INTRODUCTIONCardiovasculardiseaseisoneofthemaincausesofhu-mandeathandposesahugefactorleadingtotheriseofmedicalandhealthcarecosts.Datashowthatthenumberofpatientsworldwidewithcardiovasculardiseasesincreasedfrom271millionin1990to523millionin2019[1].Asanimportantgroupofcardiovasculardiseases,arrhythmiacancauseheartfailureandsuddendeath[2].Thus,regularmonitoringturnsouttobeparticularlyimportant.Clinically,regularmonitoringisachievedbyECGwhichcanreflecttheelectricalactivityofcardiacexcitementandiscommonlyperformedformedicalscreeningofmanycardiacdiseases,suchasdiagnosingarrhythmiaandcoronaryheartdisease[3][4].Withthedevelopmentofdeeplearningtechnique,therequirementforautomatedECGanalysisisincreasinglyrised,forallowing24-hourmonitoringandprovidingadditionaldiagnosticinformation.AlthoughlargeamountsofECGdatacanbeobtainedthroughcontinuousmonitoring,theECGannotationforsupervisedmodeltrainingislimitedintheexperiencedcardiologists,whichwillinevitablyleadtohighlaborcosts.Therefore,inordertoimprovetheclassificationperformanceofsmallsampleECGsignals,thedesignofalgorithmshasbecomeatoppriorityinpractice.ThedesignofalgorithmsforECGsmallsampleclassifi-cationshouldsatisfytheneedtomaximizetheclassificationaccuracywithlimitedsupervisedinformation.Specifically,amodelislearnedwithasmallnumberofavailabletrainingsamplestoidentifytheclassesofECGs.However,thetra-ditionalmachinelearningalgorithms[5][6][7][8]donotsolvethesmallsampleclassificationproblemwell.Themainchallengecomesfromthatthemodellosesalargeamountofinformationduringfeatureextraction,whichleadstoalowclassificationaccuracyforsmallsamples.Deeplearningalgorithmscanachieveaneffectivecombinationoffeatureextractionandclassificationthroughend-to-endlearning,thussolvingtheproblemofinformationlossinmanualfeatureextraction.However,mostofcurrentworksonlyconsidersingle-scaleinformationandignorecomplementaryinforma-tionfromotherscales[9].Asaresult,forsmallsamples,thelimitedtrainingdatadoesnotprovidesufficientsingle-scaleinformationtodemonstratethedistributionofdiseaseclasses.Howtoextractmoremeaningfulinformationfromlimitedsamplestoimprovetheclassificationperformance?InspiredbytheDMSFNetmodelproposedbyRuxinetal[10],weemployamulti-scalefeatureextractionapproachtosolvethisproblem.Multi-scalefeatureextractionnotonlyallowstoobservetheamplitudeandlocalstatisticalinformationinECGsignals,butalsoenablestobetterencodethespacinginformationbetweendifferentwavesandmorphologicalfeatures,suchasQRSduration,P-RintervalandR-Rintervaletc.Therefore,themethodofmulti-scalefeatureextractionisbeneficialtomake

管理员

195篇文档

相关搜索

B751 9145

评论

发表评论
< /7 > 付费下载 ¥ 10 元

Powered by DS文库

Copyright © 籍索 All Rights Reserved.
×
保存成功