摘要
Abstract
Objective To explore the value of a machine learning model incorporating primary tumor and peritumoral radiomics features for preoperative prediction of lymphovascular invasion(LVI)in gastric cancer.Methods Clinical and imaging data of 148 patients with pathologically confirmed gastric cancer were retrospectively collected.Based on pathological results,patients were divided into an LVI-positive group(79 cases)and an LVI-negative group(69 cases).Patients were randomly divided into a training set(103 cases)and a test set(45 cases)in a 7∶3 ratio.Radiomic features were extracted from the primary tumor and peritumoral regions.The least absolute shrinkage and selection operator(LASSO)method was used to select optimal radiomic features,and the radiomics score(Rad-score)was calculated.The clinical features with statistically significant differences between the two groups were combined with Rad-score for multivariate logistic regression analysis to select variables for constructing a machine learning model.Seven machine learning algorithms,including logistic regression(LR),extreme gradient boosting(XGBoost),random forest(RF),Gaussian naive Bayes(GNB),support vector machine(SVM),light gradient boosting machine(LightGBM),and K-nearest neighbors(KNN),were used to construct clinical-radiomics models.The performance of the models was evaluated using receiver operating characteristic(ROC)curve analysis.Calibration curves and decision curve analysis(DCA)were used to assess the calibration degree and clinical net benefit of the models,respectively.The SHapley Additive exPlanations(SHAP)method was employed to provide visual interpretation of the predictive model.Results In the training set,all seven machine learning models achieved an AUC greater than 0.650,with the RF model achieving the highest AUC(0.858),sensitivity(0.895),and accuracy(0.776).The calibration curve indicated that the RF model had the lowest Brier score(0.153),demonstrating the best predictive accuracy.DCA revealed that the RF model provided the highest net clinical benefit when the risk threshold ranged from 0.30 to 0.70.In the test set,the RF model maintained stable diagnostic performance,achieving an AUC of 0.821.SHAP analysis identified key factors associated with LVI risk in gastric cancer patients and provided visual interpretation for individual predictions.Conclusion The RF model,integrating primary tumor and peritumoral radiomic features with clinical factors,holds significant value for preoperative prediction of LVI status in gastric cancer patients.关键词
影像组学/机器学习/胃癌/脉管浸润/危险因素Key words
Radiomics/Machine learning/Gastric cancer/Lymphovascular invasion/Risk factor分类
特种医学
