基于特征解耦与融合的不完全多模态骨肿瘤图像分类

打开文本图片集

Abstract： Objective To construct a bone tumor clasification model basedon feature decoupling and fusion for processing modalityloss and fusing multimodal information toimproveclasification acuracy.Methods A decouplingcompletion module was designed toextractlocaland globalbone tumor image features from available modalities.These features were thendecomposedintosharedandmodality-specificfeatures，which wereusedtocomplete the missing modalityfeatures， therebyreducingcompletionbiascausedbymodalitydifferences.Toaddressthechallengeofmodalitydiferencesthathinder multimodalinformationfusion，acosattentinasedfusionmoulewasitroducedtoeancethemodel'sbilitytolea cros-modalinformationandfullyintegratespecificfeatures，therebyimprovingtheacuracyofbonetumorclassification. ResultsTheexperimentwasconductedusingabonetumordatasetcolectedfromtheThirdAfiliatedHospitalofSouthern MedicalUniversityfortrainingandtesting.Amongthe7availablemodalitycombinations，theproposedmethodachievedan averageAUCracydspiicityof066，.61and0793，speivelyichepresentimpoetf 2.6% 3.5% and 1.7% over existing methods for handling missing modalities.Thebest performance was observed when allthe modalities wereavailable，resulting inanAUCof 0.837，whichstillreached0.826 evenwithMRIalone.Conclusion Theproposed method can efectively handle missngmodalitiesandsuccessfully integratemultimodal information，and showrobust performance in bone tumor classification under various complex missing modality scenarios.

Keywords：bone tumor clasification; multimodalimaging; modalitymissing;feature decoupling;atentionfusion

原发性骨肿瘤（PBT）起源于骨骼或骨髓，是20岁以下癌症患者死亡的第3大原因1。（剩余13447字）