Precise T Staging of Gastric Cancer: Construction, Comparison, and Validation of Multi-Model Radiomics Based on CT Venous Phase Imaging

Objective To construct various radiomics models based on CT venous phase images for the quantitative assessment of gastric cancer (GC) invasion depth into the gastric wall, and to compare the diagnostic performance of different models in distinguishing early (T1-T2 stage) from advanced (T3-T4 stage) GC, aiming to identify the optimal model and evaluate its potential for clinical application. Methods In this retrospective study, 223 pathologically confirmed GC patients (66 early-stage, 157 advanced-stage) treated between January 2022 and May 2025 were enrolled. Patients were allocated into a training set (n = 156) and an independent test set (n = 67) through stratified random sampling at a 7:3 ratio. All patients underwent enhanced CT within one week prior to surgery. Venous phase images were selected, and three-dimensional regions of interest (ROIs) encompassing the tumor were manually delineated using 3D Slicer for radiomics feature extraction. In the training set, feature selection was performed using Least Absolute Shrinkage and Selection Operator (LASSO) regression. Four machine learning models were subsequently constructed: Logistic Regression (LR), Random Forest (RF), XGBoost (XGB), and Support Vector Machine (SVM). Model optimization was conducted via 5-fold cross-validation. Performance was evaluated using receiver operating characteristic (ROC) curves, precision-recall (PR) curves, confusion matrices, and decision curve analysis (DCA). Results A total of 148 radiomics features were extracted, from which 11 key features were selected by LASSO. In the training set under 5-fold cross-validation, the LR, RF, SVM, and XGB models demonstrated strong discriminatory ability, with AUCs of 0.912, 0.910, 0.899, and 0.880, respectively. In the independent test set, the LR model exhibited the best overall performance: AUC 0.912 (95% CI: 0.853–0.948), accuracy 0.791, sensitivity 0.787, specificity 0.800, F1-score 0.841, and average precision (AP) 0.966. DeLong's test revealed statistically significant differences in AUC among models (all P < 0.001), though the difference between LR and RF was minimal (0.0021, 95% CI: − 0.0459–0.0502). All models showed good calibration (Hosmer–Lemeshow test P > 0.05). DCA indicated that the LR model provided higher net clinical benefit within a threshold probability range of 0.3–0.8. Conclusion Among the multiple radiomics models constructed based on CT venous phase images, the LR model demonstrated the best performance in distinguishing early from advanced GC T stages, showing favorable diagnostic efficacy and clinical utility. It can serve as a quantitative auxiliary tool for preoperative staging and treatment decision-making, providing an objective basis for precise GC diagnosis and treatment.