A new artificial intelligence foundation model designed specifically for intraoperative pathology could help surgeons make faster and more accurate treatment decisions during operations, according to a study published in Nature Medicine. The model, called CRISP, is a vision-based pathology foundation model developed exclusively from frozen section slides — the rapid tissue preparations used while a patient is still on the operating table.
Unlike general pathology models trained on standard formalin-fixed, paraffin-embedded tissue, CRISP was built from the ground up using frozen section images. This distinction matters because frozen sections present unique visual characteristics, including freezing artifacts, that can trip up models trained on other types of slides. By focusing only on frozen sections, the researchers aimed to create a tool that fits directly into the surgical workflow, where time is short and decisions must be made quickly.
The study reports that CRISP outperformed current foundation models in supporting treatment decision-making throughout the surgical process. The authors describe extensive validation of the model, including testing in a prospective cohort — a stronger form of evidence than retrospective analysis alone. Prospective validation means the model was evaluated on data collected after its development, under conditions closer to real clinical use.
Intraoperative pathology plays a critical role in cancer surgery and other procedures. During an operation, a pathologist may examine frozen section slides to determine whether a tumor has been fully removed, whether lymph nodes contain cancer cells, or what type of tissue is being handled. These judgments can change the course of surgery in real time, but they are subject to time pressure and variability between observers.
A foundation model like CRISP is trained on large amounts of unlabeled data and can then be adapted to many downstream tasks. In pathology, such models have shown promise for tasks like tumor detection, classification, and grading. However, most existing models are built on standard slides, not frozen sections. The developers of CRISP argue that a dedicated frozen section model is necessary to capture the specific visual features and limitations of intraoperative imaging.
The study’s prospective cohort component is notable because many AI pathology studies rely on retrospective datasets, which can overestimate performance. Prospective evaluation helps show whether a model works in the messy, real-world conditions of an operating room, where slide quality, staining, and timing vary.
If further validated, a tool like CRISP could support pathologists and surgeons by providing rapid, consistent assessments during surgery. That could reduce the need for second surgeries, help preserve healthy tissue, and speed up decision-making. The authors position CRISP as a clinically-oriented foundation model — meaning it was designed with the surgical workflow in mind rather than as a purely research exercise.
The publication in Nature Medicine adds to a growing body of work on AI in medicine, where foundation models are increasingly being explored for their ability to generalize across tasks. For intraoperative pathology, the challenge is not just accuracy but speed and integration with existing laboratory and surgical systems. CRISP’s exclusive training on frozen sections suggests a targeted approach to those constraints.
Still, the study is a research report, and clinical adoption would require regulatory review, further multi-site validation, and careful assessment of how the model performs across different hospitals, patient populations, and cancer types. The authors’ inclusion of a prospective cohort is a step in that direction, but it does not by itself guarantee broad clinical benefit.
For now, CRISP represents a focused effort to bring foundation-model AI into one of the most time-sensitive corners of medicine: the operating room, where pathology results can shape a patient’s treatment before the surgery is even finished.
