基于CT的深度迁移学习模型对甲状腺结节BRAFV600E 突变的预测价值
打开文本图片集
[Abstract]Objective：Todevelopamachinelearningmodelbasedondeeptransferlearning（DTL）featuresextractedfrom CT images to predict the BRAFV600E mutation status in thyroid nodules.Methods ：A retrospective analysis was conducted on 256 thyroid nodules from 192 patients.According to BRAFV600E genetic testing results，the nodules were categorized into amutantgroup（117nodules）andawild-typegroup（139nodules）．Thedataset wasrandomlydividedintoatraining cohort（179nodules）anda validation cohort（77nodules）ata 7：3 ratio.The 3D ResNet-18 network was employed to extract512DTLfeaturesfromtheCTimages.Following featureselection，predictionmodelswereconstructedusing7machine learningalgorits，inludingecisiontree（D），andomforest（RF），lgisticgression（LR），supprtvctormacine（V）， extremegradientboosting（XGBoost），K-nearestneighbor（KNN），andNaiveBayes（NB）.Univariateandmultivariatelogistic regressionanalyseswereusedtoscreentheclinicalfeaturesandimagingfeaturestoconstructaclinicalmodel.Theprediction scoresofthebestmodelamong the7machine learning modelswerecombinedwiththeclinical featuresafterscreeningto constructacombinedmodel，andthenanomogramwasdrawn.Theaccuracyandclinicalutityofthecombinedmodelwere evaluatedusingcalibrationcurveanddecisioncurveanalysis.Results：SeventeenDTLfeatureswereselected.TheXGBoost modeldemonstratedthebestperformanceinthevalidationcohort，withanAUCof0.771（95%CIO.664-0.878）.Univariateand multivariatelogisticregresionanalysesshowedthat，ageandmaximumdiameterofnoduleswereindependentpredictors of BRAFvooE mutation，which wereused to construct a clinical model，with an AUC of 0.741（95%CI 0.630-0.853）in the validationcohort.ThecombinedmodelwasconstructedbythepredictionscoreofXGBoostmodel，ge，andnodulemaximum diameter，with an AUC of 0.782（95%CI 0.679-0.884）in thevalidation cohort. The calibration curve indicated good consistency between the combined model results and gene testingresults，and thedecision curveanalysisshowed that thecombined model had a favorablenet benefit.Conclusions：Thecombinedmodelutilizing the XGBoostalgorithmdemonstratsfavorablepredictiveperformanceinnoninvasively predicting the BRAFV600E mutation status in thyroid nodules，potentiallyaiding clinicians in risk assessment. [Key words] mutation;Thyroidneoplasms；Carcinoma，papilary;Thyroidnodule;Tomography，X-raycomputed; Deep transfer learning
甲状腺乳头状癌（papillary thyroid carcinoma，PTC）是最常见的甲状腺恶性肿瘤类型（占 80%～ 90%），且近年来的发病率呈显著上升趋势[1-2]BRAFV600E 突变在PTC中具有重要的诊断与预后价值[3-4]。（剩余11104字）
