融合CT影像组学、深度迁移学习和临床特征的肝细胞癌GPC3表达预测模型研究OA

Objective To investigate the application value of a combined prediction model based on CT radiomic features,deep transfer learning(DTL)features and clinical features for the non-invasive preoperative evaluation of glypican-3(GPC3)expression status in hepatocellular carcinoma(HCC).Methods Clinical and imaging data of 229 patients with pathologically confirmed HCC after surgery at a hospital from January 2016 to December 2021 were retrospectively collected.Among them,178 cases were GPC3-positive and 51 cases were GPC3-negative by immunohistochemistry.First,3D Slicer 5.7.0 software was used to delineate regions of interest(ROIs)on arterial phase and portal venous phase images of contrast-enhanced CT,and 3D volumetric ROIs of tumors were reconstructed.Second,radiomics features and DTL features were extracted from segmented ROIs of arterial and portal venous phases,and radiomics features and DTL features were further integrated to obtain deep learning radiomics(DLR)features.Finally,features were selected using t-tests or Mann-Whitney U tests,Pearson correlation analysis,minimum abso-lute shrinkage and selection operator algorithm,independent risk factors for GPC3-positive expression were identified using univariate logistic regression.Six machine learning algorithms,including Logistic regression,random forest,support vector machine,decision tree,AdaBoost and light gradient boosting machine(LightGBM),were employed to construct radiomics(Rad)model,DTL model and DLR model,respectively.The optimal model was selected and combined with clinical risk factors to establish a combined model.Receiver operating characteristic(ROC)curves were used to evaluate the predictive efficacy of the Rad model,DTL model,DLR model,clinical model and combined model for GPC3 expression status in HCC.Calibration curves and clinical decision curves were adopted to verify the consistency and practicability of each model.Delong test was performed to compare the statistical differences in area under the curve(AUC)among the models.The Shapley Additive exPlanations(SHAP)value was calculated to quantify the impact of each feature on the importance of DLR model prediction results.Results The DLR model constructed based on LightGBM was the optimal model,and alpha-fetoprotein(AFP)was an independent clinical risk factor.Compared with the other models,the combined model based on the optimal model and AFP showed the best diagnostic performance,with AUC values of 0.965 and 0.905 in the training set and test set,respectively.Calibration curves and decision curve analysis demonstrated that the established combined model had better calibration and clinical practicability than the other models.Delong test indicated statistically significant differences in AUC values between the combined model and the clinical model,Rad model,DTL model and DLR model in both the training set and test set(P<0.05).SHAP analysis revealed that wavelet_HHH_glcm_ClusterShade_P had the highest mean SHAP value(0.54),representing the most critical feature affecting model prediction outcomes.Conclusion The combined prediction model constructed on the basis of CT radiomic features,DTL features and clinical features exhibits excellent performance in predicting preoperative GPC3 expression status in HCC patients,and provides a reliable non-invasive evaluation tool for the formulation of individualized clinical treatment strategies for HCC.[Chinese Medical Equipment Journal,2026,47(4):13-24]