A GAN-based multimodal framework for preoperative response prediction in HER2-negative locally advanced gastric cancer treated with neoadjuvant chemoimmunotherapy.

111

Background: Gastric cancer is a leading global cause of cancer-related death, with approximately 77% of patients in China diagnosed at a locally advanced stage (LAGC). While neoadjuvant chemoimmunotherapy is standard, responses vary. Currently, the gold standard for evaluating treatment efficacy is postoperative pathological tumor regression grade (TRG), which cannot guide preoperative decision-making. Most existing models rely on single-modal data, such as CT imaging or pathological features, and are often limited by incomplete characterization of the tumor microenvironment or sampling bias. This study proposes a novel framework called Victory: an Incomplete Multi-modal Efficacy Prediction Model based on Generative Adversarial Networks. It imputes missing data and fuses CT and pathology features to predict treatment response preoperatively, aiming for personalized LAGC management. Methods: We analyzed 758 HER2-negative LAGC patients (2020-2025) from 3 Chinese centers. Cohorts were defined by data completeness. The Victory model was developed using a generative adversarial network to impute missing modalities, generating five types of input features. This enabled multi-modal feature fusion and prediction across various clinical scenarios.A clinical model incorporating CEA, PD-L1, tumor size, and Ca125 was also built. Model performance was evaluated using six metrics and Kaplan-Meier survival analysis. Results: The Victory model performed best with paired multimodal data. In the external validation cohort, the model trained on real paired multimodal data (rCT_rP) achieved an AUC of 0.9322, significantly outperforming models trained on single-modal data (single CT model rCT AUC = 0.6567; single pathology model rP AUC = 0.7847). Furthermore, integrating clinical data did not improve model performance. Conclusions: The Victory model establishes a novel generative artificial intelligence-based paradigm for multi-modal diagnosis and treatment. It provides an end-to-end framework—from data imputation and modality fusion to efficacy prediction—for CT imaging and pathological features. This tool offers feasible decision support for preoperative efficacy prediction and prognostic stratification in LAGC patients undergoing neoadjuvant chemoimmunotherapy, facilitating optimized individualized treatment planning.