基于胸片的迁移学习对活动性肺结核与社区获得性肺炎的鉴别诊断

打开文本图片集

[Abstract］Objective ∵ Toexplore the value of the model constructed by transfer learning based on chest X-rays in the diferential diagnosisforactivepulmonarytuberculosis（APTB）andcommunity-acquiredpneumonia（CAP）．Methods：Chest X-raysof715APTBpatientsand513CAPpatientswereretrospectivelycolectedandrandomlydividedintoatraining set（614 cases），a validation set（369 cases）and atest set（245 cases）according to the 5：3：2 ratio．Four pre-training networksincludingVGG16，Xception，ResNet5OandMobileNetwereusedfortransferlearning.Afterthetraining，themodels withthehighestacuracyofthefourpre-trainingnetworksinthevalidationsetwereselectedastheoptimalmodels，ndthe diferentitiondegree，calibrationdegreeandnetbenefitofthefouroptimal modelsinthetestsetwereevaluated.Results： Afterthetraining，VGG16andResNet5OfittdwellwhileXceptionandMobileNethadoverfitngphenomena.Themodels withthehighestaccuracyofthefournetworksinthevalidationsetwereappliedtothetestset，andtheresultsshowed thatVGG16，ResNet5O，andMobileNetoptimalmodelshadhigherAUCscomparedtoXceptionoptimalmodelandradiologist model（all P<0.05 ）.The calibration curve showed that VGGl6，ResNet5O，and MobileNet optimal' models and radiologist modelhadgoodcalibration.Decisioncurveanalysisshowedthattheoptimalmodelsofthefournetworkswerewithinalarge probabilitythresholdrange，whichmadethepatientsnetbenefit，andtheoptimalmodelofVGG16madethepatientsnet benefitthemost.Conclusions：TransferlearningtechnologyhashighclasificationperformanceinidentifyingAPTBand CAP in chest X-rays，VGG16 has the highest performance among the four kinds of pre-training networks.

[Key words] Tuberculosis，pulmonary;Community-acquired pneumonia; Convolutional neural networks

结核病死亡人数在全球传染性疾病中位居第1。（剩余8900字）