AI and the World's Hard Problems · Medicine & drug discovery · Nature (machine learning) · 2026-10-26
Comparative analysis of end-to-end learning and foundation models for clinical-grade prostate pathology
Foundation models pre-trained with self-supervision are widely promoted as a general solution for pathology, but their advantage over task-specific models is unclear. Using more than 100,000 prostate biopsies from 6,268 patients across 15 sites, the authors found that foundation models help when data is scarce, but with enough labeled data task-specific models match or exceed them. Task-specific training also markedly reduced misgrading, while foundation models consumed up to 35 times more energy.