A comprehensive computational pathology atlas for multi-cancer tumor microenvironment characterization and prognosis prediction

H&E-stained tissue specimens represent the gold standard for tumor diagnosis, yet many deep learning models for prognosis often function as opaque systems with limited clinical interpretability. To address this, we developed a transparent pipeline that histologically phenotypes slides by simultaneously segmenting 15 distinct tissue regions, including epithelium, immune infiltration, stroma, necrosis, and microvessels. We constructed a high-resolution whole-slide tissue atlas and, by integrating it with regions of interest corresponding to the tumor bulk and invasive margin, derived up to 3168 quantitative descriptors of the tumor microenvironment (TME). Across research cohorts spanning nine different tumor types, we found that these descriptors were broadly associated with patient prognosis, homologous recombination deficiency, TP53 score, tumor mutation burden, as well as diverse immune signatures and immune subtypes. By making this comprehensive atlas and paired segmentation masks openly accessible, we provide an open resource to expedite the development and validation of interpretable, H&E-based biomarkers for precision medicine.