CT–based radiomics–deep–learning fusion for predicting occult lymph node metastasis in clinically node–negative esophageal squamous cell carcinoma

Abstract Objectives To develop and validate a CT model integrating radiomic and deep–learning features for preoperative prediction of occult lymph node metastasis (OLNM) in clinically node–negative esophageal squamous cell carcinoma (cN0 ESCC). Methods This retrospective two–center study included 300 patients with cN0 ESCC who underwent surgery with lymphadenectomy. Center I patients were assigned chronologically to training ( n = 214; January 2019–December 2023) and temporal test ( n = 40; January–December 2024) cohorts; center II patients formed the external test cohort ( n = 46; January 2019–December 2023). Postoperative histopathology was the reference standard. Radiomic features were extracted from primary tumor CT images, and deep–learning features were derived from ResNet50, ResNet101, and ResNet152. Fusion models were built by combining radiomic features with ResNet–derived features. Performance was assessed using AUC, Brier score, calibration, and decision curve analysis, with SHAP used for model interpretation. Results The radiomics–ResNet101 logistic regression fusion model (R101_LR) yielded the numerically highest AUCs in the training, temporal test, and external test cohorts: 0.930 (95% CI: 0.896–0.965), 0.910 (95% CI: 0.821–0.999), and 0.886 (95% CI: 0.791–0.982), respectively. In the external test cohort, the model showed an accuracy of 0.826, sensitivity of 0.792, specificity of 0.864, and Brier score of 0.153. SHAP analysis showed contributions from both feature families, with deep–learning and radiomic features accounting for 61.7% and 38.3% of the total mean absolute SHAP contribution, respectively. Conclusion The CT–based radiomics–deep–learning fusion model showed favorable discrimination for preoperative prediction of OLNM in cN0 ESCC across temporal and external testing. The model may provide complementary information for individualized preoperative nodal risk assessment, although prospective multicenter validation is required before its clinical utility can be established.