Identifying Advanced Esophageal Squamous Cell Carcinoma Patients Who Benefit from Immunotherapy Maintenance Following Immunotherapy and Radiotherapy: A Multicenter Study

Objective Patients with advanced esophageal squamous cell carcinoma (ESCC) who have received first-line immunotherapy and radiotherapy remain uncertain as to which individuals may benefit from immunotherapy maintenance. This study aims to establish a deep learning model based on interactive automated 3D segmentation to predict immunotherapy + radiotherapy efficacy and identify patients likely to benefit from immunotherapy maintenance. Methods This study collected CT images and clinical information from three centers for patients with advanced ESCC. This study employs the validated high-precision, strong-generalization-capability VISTA3D to process CT images. By combining automatic segmentation with point-prompt segmentation guidance, it achieves automated tumor identification in ESCC patients and establishes a DL model based on ResNet18. The model's performance is evaluated using metrics such as AUC. Based on this model, patients who may benefit from immunotherapy maintenance are identified. Results The study included a total of 362 patients. When performing automatic segmentation based on the VISTA3D segmentation model, the model achieved a Dice coefficient of 0.87 on the external validation cohort, with a C-index reaching 0.86 (95% CI: 0.73–0.94). In analyses of PFS and OS, patients classified as low risk by the model demonstrated significantly better outcomes than those classified as high risk (HR: 0.61, 95% CI: 0.46–0.79, P < 0.001; HR: 0.55, 95% CI: 0.39–0.77, P < 0.001). Among high-risk patients, those receiving immunotherapy maintenance demonstrated significantly improved PFS and OS (HR: 0.39, 95% CI: 0.26–0.75, P < 0.001; HR: 0.57, 95% CI: 0.31–0.89, P = 0.023). Conclusion The DL model developed in this study, based on automated segmentation combined with point-prompt segmentation, can replace prediction models relying on manual segmentation. This model effectively assists clinicians in predicting the prognosis of advanced ESCC patients undergoing immunotherapy and radiotherapy, validating the feasibility of a two-stage automated prediction framework integrating “image segmentation-prognosis prediction”. Additionally, patients identified as high-risk by the model can significantly benefit from immunotherapy maintenance. This approach supports clinical decision-making for advanced ESCC patients, helps avoid overtreatment, and advances the progress toward personalized and precision cancer care.