Harnessing foundation models for digital pathology without re-training

Researchers have developed PRET, a training-free framework that adapts pathology foundation models to a wide range of clinical oncology tasks without the need for labeled data or model retraining. Published in Nature Cancer, the study demonstrates robust performance across pan-cancer diagnosis applications including cancer screening, tumor subtyping, tissue segmentation, and metastasis detection. The approach works entirely at the inference stage, eliminating one of the most significant bottlenecks in deploying AI-based digital pathology — the costly and time-consuming process of curating labeled datasets and retraining models for each new task. If validated in prospective clinical settings, PRET could substantially accelerate the adoption of AI-powered pathology tools across diverse cancer types and healthcare systems.