Development and Validation of an AI Foundation Model for Frozen-Section Pathology

Frozen-section pathology is an essential component of intraoperative consultation and provides rapid diagnostic information that may support surgical decision-making. However, frozen-section images differ substantially from formalin-fixed, paraffin-embedded tissue sections because of freezing artifacts, variable section quality, staining variation, and differences in tissue morphology. Most currently available pathology foundation models have been developed primarily using formalin-fixed, paraffin-embedded pathology images and may have limited generalizability to frozen-section pathology.

This multicenter, observational study is designed to develop and validate an artificial intelligence foundation model for frozen-section pathology. The study consists of a retrospective phase and a prospective validation phase.

In the retrospective phase, patients with frozen-section pathology data collected from January 2019 to April 2026 will be included for model development, internal validation, and external validation. The retrospective development and internal validation datasets will include patients from Sun Yat-sen Memorial Hospital, Sun Yat-sen University Cancer Center, and the Third Affiliated Hospital of Sun Yat-sen University. Retrospective external validation will include patients from collaborating hospitals, including Shantou Central Hospital, the First Affiliated Hospital of Shantou University Medical College, and the Fifth Affiliated Hospital of Sun Yat-sen University.

In the prospective phase, patients undergoing intraoperative frozen-section examination will be enrolled from June 2026 to October 2026, including patients from Sun Yat-sen Memorial Hospital and patients from the Fifth Affiliated Hospital of Sun Yat-sen University. The study population will include patients with benign or malignant diseases involving the lung, thyroid, stomach, colorectum, liver, bladder, prostate, kidney, uterus, and ovary.

For each eligible patient, clinical information, frozen-section whole-slide images, intraoperative pathology reports, and corresponding pathological data will be collected. Whole-slide images will undergo tissue detection, region-of-interest segmentation, and patch-level image processing. The model will be pretrained using self-supervised learning and subsequently fine-tuned for prespecified downstream frozen-section pathology tasks using multiple-instance learning or Transformer-based aggregation methods. Model interpretability will be explored using gradient-weighted class activation mapping.

The primary outcome will be the discriminative performance of the artificial intelligence foundation model, assessed using the area under the receiver operating characteristic curve. Secondary outcomes will include accuracy, sensitivity, specificity, positive predictive value, negative predictive value, and robustness across participating centers, organ systems, and image-quality conditions.

Inclusion criteria

• Patients undergoing surgery with intraoperative frozen-section pathological examination.
• Availability of complete clinical information and intraoperative frozen-section pathology records.
• Availability of digitized frozen-section whole-slide images suitable for artificial intelligence analysis.

Exclusion criteria

• Frozen-section whole-slide images with inadequate quality for evaluation, including substantial blur, ghosting, severe artifacts, or insufficient diagnostic tissue.
• Missing or indeterminate key clinical, intraoperative pathology, or pathological reference data required for the prespecified study task.
• Withdrawal of informed consent in the prospective validation cohort, where applicable.