CT radiomics of adenocarcinoma of the esophagogastric junction: machine learning-based prediction of perineural invasion status

Preoperative diagnosis of perineural invasion (PNI), a critical prognostic factor in adenocarcinoma of the esophagogastric junction (AEG), remains challenging. This study aimed to develop and validate a computed tomography (CT)-based radiomics model for the noninvasive prediction of PNI and to evaluate the incremental value of integrating radiomics with clinical risk factors. A retrospective cohort of 306 patients with pathologically confirmed AEG was randomized into training (n = 214) and testing (n = 92) cohorts. Radiomics features were extracted from venous-phase CT images using PyRadiomics. Following feature selection with the least absolute shrinkage and selection operator (LASSO) algorithm, five machine learning classifiers—logistic regression (LR), random forest (RF), extra trees (ET), adaptive boosting (AdaBoost), and support vector machine (SVM)—were trained to build the radiomics model. A clinical model was constructed using independent risk factors identified by multivariate logistic regression, which were then integrated with the radiomics signature to build a combined clinical-radiomics model. Model performance was assessed using receiver operating characteristic (ROC) analysis, calibration curves, decision curve analysis (DCA), and standard classification metrics. Among the patients, 224 (73.2%) were PNI-positive. The radiomics model based on RF demonstrated the best performance, with an area under the curve (AUC) of 0.878 in the training cohort and 0.808 in the testing cohort. The clinical model achieved lower AUCs (0.747 and 0.636, respectively). The combined model significantly outperformed both, achieving AUCs of 0.932 and 0.861 in the training and testing cohorts (all P < 0.05, DeLong test), with good calibration and the highest net clinical benefit on DCA. A CT-based clinical-radiomics model can effectively predict PNI status in AEG preoperatively. The integration of radiomic features with clinical parameters further enhances predictive performance and clinical utility, offering a valuable noninvasive tool to guide individualized treatment planning.