Researchers have built CRISP, a pathology foundation model trained exclusively on frozen sections, the thin slivers of tissue pathologists examine while a patient is still on the operating table. The work appears in Nature Medicine.
Surgeons lean on frozen sections to decide how much tissue to take. Turnaround is measured in minutes, and the data available for training models is scarce.
CRISP was built to close that gap. Its developers gathered 100,000-plus frozen sections from ten medical centers. Evaluation then covered more than 15,000 intraoperative slides and close to 100 diagnostic tasks.
The model generalized across six institutions, 14 tumor types and 24 anatomical sites, including sites and rare cancers it had never seen. In a prospective cohort of more than 3,000 patients it held high accuracy under real-world conditions and informed surgical decisions in 92.6% of cases.
Pairing pathologists with the model cut diagnostic workload by 35%, avoided 105 ancillary tests and detected micrometastases with 87.5% accuracy. Those gains are the point: computational pathology has produced strong retrospective benchmarks but very little prospective evidence that it changes care.
