CRISP

A frozen section is cut, stained and read while the patient is still on the operating table, usually within twenty minutes. That speed is paid for in tissue quality: freezing leaves ice-crystal holes, the sections are thicker and more compressed than a formalin-fixed paraffin-embedded (FFPE) slide, and nuclear detail is smeared. Nearly every pathology foundation model of the past few years was pretrained on FFPE archives — the one preparation an intraoperative pathologist never sees — so applying those representations to a frozen section means asking them to generalize across a domain gap they were never shown. CRISP, for Clinically-oriented Robust Intraoperative Support for Pathology, was built to remove that mismatch by pretraining on frozen sections directly.

The model comes from Sun Yat-sen University Cancer Center and the Hong Kong University of Science and Technology, with pathology departments at seven further Guangdong hospitals contributing cohorts. It was posted as a preprint in October 2025 and published in Nature Medicine on 10 September 2026.

Rather than train a new backbone, the authors adapt Virchow2, a vision transformer pretrained on roughly a million FFPE whole-slide images, using low-rank adaptation (LoRA). Only the injected rank-decomposition matrices are updated, so the frozen-section signal is learned without overwriting the general morphological knowledge already in the backbone. It is the same subspecialty-adaptation recipe the group applied to lung histology in PulmoFoundation, pointed here at a preparation artifact rather than an organ system.

#Key Features

• Frozen-section-native pretraining: Self-supervised pretraining runs on over 100,000 frozen-section whole-slide images gathered from ten medical centres, the largest such corpus assembled for this preparation.
• Parameter-efficient domain adaptation: LoRA modules on the query and value projections of each attention layer train only about 2 million parameters, 0.3% of the model's total, which keeps the adaptation cheap and limits catastrophic forgetting.
• One checkpoint, ~100 tasks: A single fixed encoder supports benign–malignant discrimination, intraoperative decision-making and pan-cancer detection, with slide-level predictions produced by attention-based multiple-instance learning over the frozen features.
• Transfer to unseen sites: Generalization holds across 6 institutions, 14 tumour types and 24 anatomical sites, including sites and rare cancers absent from pretraining.
• Prospective clinical evaluation: A registered observational study, rather than a retrospective split, measures how the model behaves inside a live surgical workflow.

#Technical Details

CRISP is a ViT-H/14 encoder with 633 million parameters in total, inherited from Virchow2 and adapted with LoRA at rank 8 and scaling factor 16; the released checkpoint emits a 2,560-dimensional patch embedding for 224×224-pixel inputs. Adaptation uses a DINO self-supervised objective over roughly 50 million image patches drawn from 25 anatomical sites, with pretraining slides and evaluation slides kept disjoint at the cohort level rather than split within a cohort. Retrospective evaluation covers more than 15,000 intraoperative slides across nearly 100 diagnostic tasks, with external validation cohorts drawn from institutions that contributed no pretraining data. In the prospective arm — over 3,000 patients — CRISP's output directly informed the surgical decision in 92.6% of cases; pathologist–model collaboration cut diagnostic workload by 35%, avoided 105 ancillary tests, and detected micrometastases at 87.5% accuracy.

#Applications

The intended user is a pathologist on intraoperative call, where the question is usually narrow and the clock is the binding constraint: is this margin involved, is this node positive, is this lesion malignant enough to change the planned resection. CRISP supplies frozen features that a lightweight classifier or attention aggregator can be fitted to per task, which makes it practical for a hospital to stand up a local model for its own case mix without pretraining anything. The reported workload and ancillary-test reductions point at the second use — triage, where the model handles routine calls and flags the cases that warrant a senior read or a deferral to permanent sections.

#Impact

Frozen-section support has been a conspicuous gap in computational pathology: the clinical need is acute and the data are scarce, because frozen sections are diagnostic by-products rather than archived specimens. CRISP shows that the gap can be closed by adaptation rather than by another from-scratch pretraining run, and that a few million trainable parameters on top of an existing FFPE backbone are enough. Its prospective arm is the more consequential contribution — most pathology foundation models are reported on retrospective benchmarks alone, and this one is evaluated against decisions surgeons actually acted on. The evaluation remains single-health-system for the prospective component and Guangdong-wide for the retrospective cohorts, so performance in other populations and scanner estates is not yet established, and the model is a research artifact without regulatory clearance for diagnostic use.