基于多級多特征混合模型的白血病亞型自動分類
打開文本圖片集
中圖分類號：TP391 文獻(xiàn)標(biāo)志碼：A 文章編號：1001-3695（2025）06-038-1880-07
doi：10.19734/j.issn.1001-3695.2024.08.0354
Automatic classification of leukemia subtypes based on multi-scale multi-feature hybrid model
Gao Mingyang1，Geng Yan2，Yu Xiao3，Pei Bo4，Zhao Juanjuan1，Qiang Yan1，5t（1.ColegeofmuterSece &Tooy（lgeofatSce），anUiestofogin;2.SchoolfsieicalciediclUesitinasspialfieclUTaiyuanO3Uesfd6UfeUf030051，China）
Abstract：Leukemia，ahighlyconcealedcancer，presentssignificantchalengesinearlydetection，makingitafocalpointfor medical professionals.Existing fine-grained clasificationmodelsstrugglewithsmallsampleimbalanceddatasets，leadingto poorperformance in classfying leukemiasubtypes.Toaddressthese issuesandacceleratedoctors’diagnosticspeed while shortening treatment time，this paperproposedamulti-scalemulti-feature hybrid model（MSMFHM）fortheautomaticclassificationofleukemiasubtypesappicable tosmallsampledatasets.The modelfirstly extractedmulti-levelstructural features from imagesusingamulti-scalefeature extractionframeworkcombined withscalingoperationsandaCNNbackbone.Next，amultiscalefusion modulewith atention mechanisms integratedthesemulti-level structural featuresandextracted fine-grained features，effectivelyleveraging therobustnessofCNNinductivebiasandthecomplexglobalmodelingcapabilitiesofTransformers.Finally，toenhance robustnessand mitigate overfiting issues causedbysmallsamples，amulti-feature hybrid module combined texturefeatures withfine-grained features beforeclasification.Adatasetof7156leukemiacellimages，along with otherrelevantpublicdatasets，wascollctedtoevaluatethismethod.Theproposedmodelachievesclasificationaccuraciesof （204號 93.03% and 99.42% on private and public datasets，respectively，outperforming other advanced models. This method accuratelydistinguishedacute leukemia subtypecellsandservesasanefectivedesignapproach forcomputer-aideddiagnosisof leukemia.
Key words：leukemia；medical image processing；multi-scale feature fusion； Transformer
0 引言
全球造成成千上萬人死亡[1]。（剩余17275字）