Abstract / Summary
Abstract Personalized management of colorectal cancer (CRC) requires accurate risk stratification and identification of clinically actionable molecular alterations. However, standard clinicopathologic features often fail to capture the cellular and tumor microenvironmental heterogeneity driving disease behavior. Although routine hematoxylin and eosin (H&E) whole-slide images (WSIs) reflect these patterns, patch-level foundation models may obscure fine-grained cellular morphology and spatial organization. We present PathVista , a cell-centered framework that integrates cellular and tissue-level representations for prognostic and molecular stratification. PathVista incorporates CellSwAV , a cellular foundation model pretrained on more than 160 million CRC cells and combines spatially aggregated cellular features with broader tissue context. Across 2,568 CRC patients , PathVista risk scores were significantly associated with overall survival (HR = 6.65, 95% CI: 3.53–12.51) , disease-free survival (HR = 3.15, 95% CI: 1.82–5.45) , and recurrence-free survival (HR = 3.04, 95% CI: 1.76–5.26) , with risk scores remaining independently associated with outcomes after clinicopathologic adjustment. Locked models also generalized to external TCGA (HR = 5.25, 95% CI: 2.29–12.05) and SURGEN (HR = 2.89, 95% CI: 1.86–4.49) cohorts. PathVista predicted microsatellite instability (AUC = 0.91) and BRAF mutation status (AUC = 0.83) . Across 4,667 patients spanning CRC, lung, and breast cancers, PathVista demonstrated the potential of cell-centered histopathology for prognostic, molecular, and tumor characterization.