Comparative analysis of end-to-end learning and foundation models for clinical-grade prostate pathology

Foundation models (FMs) using self-supervised pre-training are widely advocated as a universal solution for pathology, yet their advantages over task-specific (TS) models remain unclear. Here we show, using over 100,000 prostate biopsies from 6268 patients across 15 sites, that while FMs demonstrate utility in data-scarce scenarios, their performance converges with—and is in some cases surpassed by—TS models with sufficient labeled training data. Concordance with pathologists’ grading reaches quadratic Cohen’s kappas of 0.91 (95% CI 0.90-0.92; all models) on internal and 0.87 (0.86-0.88; TS), 0.86 (0.84-0.87; UNI FM) and 0.85 (0.84-0.86; Virchow2 FM) on external cohorts at full training. Task-specific training markedly reduces misgrading, misdiagnosis of challenging morphologies, and technical variability. Notably, FMs consume up to 35 times more energy. We advocate for considering both FMs and end-to-end learning for robust, resource-efficient, and clinically applicable AI pathology.