Longitudinal CT-based deep learning radiomics for predicting prognosis in esophageal squamous cell carcinoma treated with definitive chemoradiotherapy: a two-center study.

Purpose: To develop and validate a longitudinal deep-learning based survival prediction model for esophageal squamous cell carcinoma (ESCC) patients who received definitive concurrent chemoradiotherapy (CCRT). Methods: A total of 257 ESCC patients from two centers were recruited. Among them, 205 patients were in the training cohort and 52 in the external testing cohort. The CrossFormer algorithm was utilized to extract features from pre- and post- treatment CECT scans. We constructed clinical, Delta-radiomics and deep-learning models. Models were evaluated by the C-index and integrated Brier score (iBS). Prognostic stratification was performed based on risk scores, and model interpretability was evaluated using Grad-CAM and SurvSHAP(t). Results: The Fusion model demonstrated superior predictive performance, achieving a C-index of 0.768 (95% CI: 0.731-0.804) in the training cohort and 0.734 (95% CI:0.665-0.803) in the testing cohort. The Fusion model also showed the best calibration (Training cohort: iBS=0.096; Testing cohort: iBS=0.151). Patients were stratified into high-risk and low-risk groups based on risk scores, with significant differences in overall survival (OS) between the groups ( Conclusion: We developed a model integrating longitudinal CECT scans to predict OS in ESCC patients undergoing CCRT. The results highlight the importance of capturing tumor changes during treatment for accurate prognostic stratification. The model shows its potential for guiding personalized treatment strategies in clinical practice.