Morphology-Aware Prognostic Model for Five-Year Survival Prediction in Colorectal Cancer from H&E Whole-Slide Images: A Study Using Multi-Center Clinical Trial Cohort

Simple Summary Colorectal cancer (CRC) poses a significant global health burden, with over 154,000 new cases projected for 2025. Accurate survival prediction remains a critical challenge, as current pathological assessments fail to capture the full spectrum of morphological heterogeneity relevant to patient outcomes. This study presents PRISM (Prognostic Representation of Integrated Spatial Morphology), a novel AI model designed for five-year overall survival prediction in stage III CRC from routine H&E whole-slide images. PRISM integrates morphology-informed features derived from expert-annotated tissue regions with foundation model embeddings, enabling a biologically meaningful representation. PRISM is trained on over 15 million histological patches from 2957 patients, and achieved an AUC of 0.70 and an accuracy of 68.37%, outperforming existing CRC-specific and general-purpose AI models. We also demonstrated robust performance across sex, treatment regimen, and clinicopathological subgroups, representing a step toward interpretable AI-driven prognostication in CRC.