A Multimodal Deep Learning Framework Integrating Tumor Microenvironment Niches for Overall Survival Prediction in Esophageal Squamous Cell Carcinoma

Accurate prediction of overall survival (OS) in esophageal squamous cell carcinoma (ESCC) is hindered by the limited prognostic resolution of conventional staging. Here, we developed MENISA, a multimodal deep learning framework integrating clinical data, CT imaging, whole‐slide pathology images, and annotated tumor microenvironment (TME) niches, specifically perineural and lymphovascular invasion to predict OS. In two multicenter cohorts totaling 450 patients, the model achieved robust predictive performance (internal validation C‐index: 0.778; external validation C‐index: 0.675) and significantly stratified patients into distinct risk groups ( p ≤ 0.001). Interpretability analysis identified TME niches as important prognostic drivers, bridging the gap between macroscopic radiomics and mesoscopic histology by introducing intermediate‐scale TME niches as a complementary prognostic dimension. This study demonstrates the clinical utility of integrating intermediate‐scale microenvironmental features, offering a precision oncology tool for personalized risk stratification in ESCC.