Preoperative prediction of HER2 expression in gastric cancer based on different machine learning models based on venous-phase contrast-enhanced CT imaging

LI Yongliang; WANG Cuiyan

doi:10.20258/j.cnki.1006-9011.2025.09.020

Abstract：Objective To explore the predictive value of seven machine learning models based on venous-phase contrast-enhanced CT radiomics features before surgery for the expression of human epidermal growth factor receptor 2(HER2)in gastric cancer.Methods A total of 119 patients with gastric cancer were selected and randomly divided into a training set of 83 cases and a test set of 36 cases at a ratio of 7∶3.CT scans of the upper abdomen or the entire abdomen were performed before surgery,and the venous-phase CT images were obtained for the study.The 3D Slicer software was used to perform three-dimensional delineation of the tumor in the venous phase images,and the Pyradiomics package was used to extract radiomics features.The intraclass corre-lation coefficient(ICC),Pearson correlation analysis,and the minimum absolute shrinkage and selection operator(LASSO)algo-rithm were used for dimensionality reduction and screening of radiomics features.Then,seven machine learning classifiers includ-ing decision tree(DT),K-nearest neighbor(KNN),Linear Support Vector machine(Linear SVC),Extreme Gradient Boosting tree(XGBoost),Random Forest(RF),Support Vector Machine(SVM),and Logistic regression were respectively utilized to con-struct predictive models for radiomics features with the random splitting method.The predictive efficacy of each model was evalu-ated using the receiver operating characteristic(ROC)curve and its area under the curve(AUC).Results Based on the venous-phase contrast-enhanced CT images,129 radiomics features were extracted.Feature screening and dimensionality reduc-tion were carried out through ICC,Pearson correlation analysis and LASSO algorithm,and finally,33 optimal radiomics feature subsets were retained.In the training set based on DT,KNN,Linear SVC,XGBoost,RF,SVM,Logistic,the AUC of the seven machine learning models of Regression for preoperative prediction of HER2 expression in gastric cancer were 0.912,0.922,0.966,0.968,0.961,0.941,and 0.947,respectively;while,in the test set,the AUC was 0.665,0.720,0.809,0.732,0.780,0.874,and 0.776,respectively.There was no statistically significant difference in the predictive performance of the seven ma-chine learning models in pairwise comparisons between the training set and the test set(all P>0.05).The calibration curves of the test set showed that the theoretical and actual curves of all seven machine learning models closely approximated the reference line.Conclusion All seven machine learning models constructed from the preoperative venous-phase contrast-enhanced CT radiomics features demonstrated high predictive efficacy for HER2 expression in gastric cancer.Among the seven machine learning models,the SVM model was proved to have the highest predictive performance for HER2 expression in gastric cancer.