Real-time tissue diagnosis during surgery has long been a bottleneck where speed and accuracy are in direct tension — pathologists must issue rapid verdicts on frozen tissue sections while surgeons wait, and errors can alter a patient's operative course irreversibly. A vision-based AI foundation model purpose-built for this exact context represents a meaningful departure from general-purpose pathology AI, which is typically trained on permanent, formalin-fixed sections that look substantially different from frozen tissue.
The model, called CRISP, was developed exclusively on frozen section slides — the actual intraoperative specimen type — rather than adapted from permanent-section datasets. Trained and validated across a prospective cohort in addition to retrospective data, CRISP demonstrated superior classification and diagnostic performance compared to existing pathology foundation models on tasks directly relevant to surgical decision-making: margin assessment, tumor identification, and subtype differentiation. The prospective validation component is particularly notable, as most AI pathology tools are evaluated only on historical datasets, which can mask real-world performance gaps.
This work sits at the intersection of two rapidly evolving fields: computational pathology and intraoperative decision support. Most foundation models in digital pathology — including large-scale efforts like UNI and Prov-GigaPath — are optimized for downstream analysis of permanent sections in diagnostic workflows, not for the compressed, artifact-laden conditions of frozen section interpretation. CRISP's domain-specific training strategy mirrors a broader lesson emerging across medical AI: general pretraining benefits from fine-tuning on the precise imaging modality and clinical moment where the tool will actually be deployed. Key limitations worth acknowledging include the single-institution or limited-center nature of most intraoperative AI studies, potential variation in frozen section quality across pathology labs globally, and the model's performance on rare tumor subtypes. Nonetheless, prospective validation elevates this beyond incremental work — it is one of the more clinically grounded AI pathology contributions of recent years.