Sheng Chen, Ping’an Ding, Yang Yu, Shuo Ma, Honghai Guo, Xiao Han, J. Joshua Yang, Wenqian Ma, Ning Meng, Zhijia Xia, Xiaolong Li, Lilong Zhang, Yanlong Shi, Zhenjiang Guo, Kaixuan Gao, Renjun Gu, Hong Long, Lingjiao Meng, Q. Zhao
Gastric cancer staging is frequently limited by the low sensitivity of routine imaging for occult peritoneal metastasis (OPM), necessitating invasive staging laparoscopy. We developed a Multimodal Model, integrating primary tumor radiomics from CT with clinical factors to non-invasively predict OPM in locally advanced gastric cancer. The model was trained and internally validated in a large cohort (n = 940) and externally validated across two independent multi-center cohorts (n = 309), an incremental cohort (n = 477), and a prospective clinical trial cohort (n = 168). In all cohorts, the model achieved robust performance (AUCs: 0.834-0.857), significantly outperforming single-modality models. Crossover validation showed AI assistance increased the average radiologist AUC from 0.735 to 0.872. Transcriptomic analysis revealed that the model's low-risk stratification correlated with an enhanced antitumor immune microenvironment (CD8 T cells, TNFα signaling). This validated model provides a practical tool for accurate, non-invasive OPM prediction and individualized treatment planning.