Clinically informed intermediate reasoning enables generalizable prostate cancer prognostication through machine...

Machine learning has shown promise in medical image classification. However, its generalizability remains challenging. Here, we show that data-efficient pre-surgical prognostication of prostate cancer from biopsy specimens is enabled by versatile feature extraction from whole-mount histopathology and a clinically informed intermediate reasoning step. With data from multiple institutions, our pipeline resolved dual-domain shifts across specimen types and institutions and achieved consistent external validation, reinforced by comprehensive analyses of generalizability. This highlights the robustness of our prognostic approach when compared to the Gleason grading system. We establish an equitable, interpretable, and clinically applicable framework, supporting actionable decisions for prognosis and treatment planning, even in limited real-world clinical environments.