A CT-based deep learning model with biological interpretation for predicting early recurrence after neoadjuvant therapy in gastric cancer.

Objective: To achieve accurate prediction of early recurrence (ER) in locally advanced gastric cancer (LAGC) patients after neoadjuvant therapy (NAT) and surgery, we constructed a deep learning fusion model integrating preoperative computed tomography (CT) imaging and perioperative clinicopathological features. Methods: We retrospectively enrolled 611 LAGC patients who received NAT from four tertiary teaching hospitals, including a training cohort (TC), an internal validation cohort (IVC), and an external validation cohort (EVC). ER was defined as recurrence occurring within 2 years post-surgery. Based on preoperative 2.5D CT images, we constructed a deep learning signature (DLS) using a ResNet50 architecture. In parallel, a clinical signature (CLIS) was developed through logistic regression analyses. To further improve predictive performance, a deep learning fusion signature (DLFS) was constructed by integrating the DLS and CLIS. Model performance was evaluated by discrimination, calibration, and clinical utility. Kaplan-Meier analysis was used to evaluate prognostic differences across risk groups. Bulk and single-cell transcriptomic analyses explored biological features. Results: Compared with the DLS and CLIS, the DLFS demonstrated superior performance in predicting ER, with area under the curve (AUC) values of 0.884 in the TC, 0.828 in the IVC, and 0.748 in the EVC. Calibration curves exhibited good agreement, and decision curve analysis indicated a higher net benefit. Risk stratification based on DLFS showed worse overall survival (OS) in the high-risk group (3-year OS: TC, 35.36% Conclusions: The DLFS model, integrating preoperative CT imaging and perioperative clinicopathological variables, effectively predicts ER and survival in LAGC patients after NAT. Thus, it can serve as a useful tool to optimize prognostic monitoring.