Interpretable deep learning maps spatial morphological phenotypes to HCC and LUAD survival

The clinical adoption of deep learning in pathology is largely hindered by the "black box" nature of current models, which often obscure the biological basis of prognosis. Here, we introduce MorphX, an interpretable framework that predicts patient survival by explicitly decoding multi-scale morphological patterns from whole-slide images. Unlike conventional methods relying on latent features, MorphX fuses macro-scale tissue texture with micro-scale nuclear metrics to attribute risk to interpretable biological phenotypes. In a multi-center study spanning 1,067 patients with hepatocellular carcinoma (HCC) and lung adenocarcinoma (LUAD), MorphX achieves concordance indices of 0.775-0.799 in discovery cohorts. Crucially, we derive eight distinct morphology-based biomarkers-driven by texture heterogeneity in necrotic regions, lymphocyte density, and nuclear shape-that demonstrate reproducible survival associations across independent cohorts, offering a granular resolution of tumor aggressiveness that transcends traditional histological grading. By translating opaque computational predictions into validated biological insights, MorphX supports trustworthy, transparent, artificial intelligence (AI)-driven precision oncology.