Accuracy of Foundation Artificial Intelligence Models for Hepatic Macrovesicular Steatosis Quantification in Frozen...

Accurate intraoperative assessment of macrovesicular steatosis in donor liver biopsies is critical for transplant decisions but is often limited by interobserver variability and freezing artifacts that can obscure histologic details. Artificial intelligence (AI) offers a potential solution for standardized and reproducible evaluation.

To evaluate the diagnostic performance of 2 self-supervised learning (SSL)-based foundation models, Prov-GigaPath and UNI, for classifying macrovesicular steatosis on frozen liver biopsy sections, compared with assessments by surgical pathologists.

This retrospective study included 131 frozen liver biopsy specimens from 68 donors collected between November 2022 and September 2024. Slides were digitized into whole slide images, tiled into patches, and used to extract embeddings with Prov-GigaPath and UNI; slide-level classifiers were then trained and tested. Intraoperative diagnoses by on-call surgical pathologists were compared with ground truth determined from independent reviews of permanent sections by 2 liver pathologists. Accuracy was evaluated for both a 5-category classification and a clinically significant binary threshold (<30% versus ≥30%).

For the binary classification, Prov-GigaPath achieved 96.4% accuracy, UNI 85.7%, and surgical pathologists, 89.3% (P = .37). For the 5-category classification, accuracies were lower: Prov-GigaPath, 57.1%; UNI, 50.0%; and pathologists, 64.2% (P = .47). Misclassification occurred mainly in intermediate categories (5% to <30% steatosis).

SSL-based foundation models performed comparably to surgical pathologists at the clinically relevant threshold of less than 30% versus 30% or greater. These findings support the potential role of AI in standardizing intraoperative evaluation of donor liver biopsies; however, the small sample size limits generalizability and requires validation in larger, balanced cohorts.