Summary
A Nature Medicine study reports CRISP, a clinically oriented pathology foundation model developed using more than 100,000 frozen sections, with prospective evaluation in more than 3,000 patients.
A clinically oriented artificial-intelligence model for pathology supported surgical decision-making in a prospective cohort of more than 3,000 patients, according to a study published in Nature Medicine on 10 September 2026.
The model, called CRISP, was developed using more than 100,000 frozen sections from 10 medical centers. Intraoperative pathology examines tissue while an operation is underway, helping clinicians assess disease and make decisions during surgery. Frozen-section analysis involves rapidly freezing tissue, preparing thin sections and examining them under a microscope while the procedure continues.
The study reports that CRISP directly informed surgical decisions in 92.6% of cases in the prospective cohort. The authors also report that collaboration between the model and human diagnosticians reduced diagnostic workload by 35%, avoided 105 ancillary tests and detected micrometastases with 87.5% accuracy.
A model built around frozen-section pathology
CRISP is described as a foundation model designed to provide Clinically-oriented Robust Intraoperative Support for Pathology. In this context, a foundation model is a system developed from a large and varied dataset for use across multiple related tasks rather than a single narrowly defined classification problem.
The researchers evaluated CRISP on more than 15,000 intraoperative slides across nearly 100 retrospective diagnostic tasks. These included distinguishing benign from malignant tissue, supporting key intraoperative decisions and detecting cancer across multiple tumour types.
The model showed what the paper describes as robust generalisation across six institutions, 14 tumour types and 24 anatomical sites. The evaluation included anatomical sites not represented previously and rare cancers, according to the study. This breadth is important because pathology systems can encounter substantial variation in tissue appearance, disease type and clinical setting.
Prospective assessment during surgery
The prospective cohort tested CRISP under what the authors describe as real-world conditions rather than only on previously collected slides. In that cohort, the model maintained high diagnostic accuracy and directly informed surgical decisions in 92.6% of cases.
The workflow findings connect model performance with the practical demands of intraoperative pathology. A 35% reduction in diagnostic workload could help pathologists handle time-sensitive examinations, while avoiding 105 ancillary tests indicates that the reported collaboration affected the use of additional diagnostic procedures. The study also reports 87.5% accuracy for detecting micrometastases, which are very small deposits of tumour cells that may be difficult to identify during routine examination.
Taken together, the findings position CRISP as a support system for pathology during surgery rather than as a model evaluated only on an isolated image-recognition task. The paper’s central evidence combines model development, retrospective testing across many diagnostic tasks and a prospective cohort assessment.
The abstract reports high diagnostic accuracy for the prospective cohort but does not provide one overall accuracy value or task-by-task performance figures. It also reports diagnostic and workflow outcomes rather than postoperative patient outcomes or results from a randomized comparison. Nature Medicine identifies the publication as an early peer-reviewed accepted version that is subject to further edits before the final Version of Record.