A deep learning model for intraoperative diagnosis of lung adenocarcinoma using frozen section whole-slide images (CryoPath): A multicentric diagnostic study.

8024 Background: Intraoperative frozen section (FS) analysis is critical for guiding surgical decision-making in lung cancer. However, its diagnostic accuracy can be limited by time constraints, sampling issues, and inter-pathologist variability. Deep learning-based analysis of whole-slide images (WSIs) offers a new paradigm for real-time, objective intraoperative diagnosis. Methods: We developed and validated an artificial intelligence system, CryoPath, for the intraoperative diagnosis of lung adenocarcinoma (LUAD) from FS-WSIs. The system was trained and tested using a large, multicentric, retrospectively collected cohort comprising 3000 LUAD FS-WSI cases from seven medical centers in China. The model architecture is based on attention-challenge multiple instance learning framework with vision transformer for feature extraction. CryoPath was designed for end-to-end prediction of diagnostic subtypes (pre-invasive lesions, minimally invasive adenocarcinoma and invasive adenocarcinoma) and key pathological features such as extent of invasion and lymphovascular invasion from FS-WSIs. Results: The study included 1,042 patients, yielding 1,139 WSI with early-stage lung adenocarcinoma from multiple centers, which comprised a total of 637 WSIs with IAC, 402 with MIA, and 100 with PIL. The CryoPath model achieved an AUC of 0.931 for discriminating IAC from PIL/MIA, which achieved a sensitivity of 0.87 and a specificity of 0.85. In clinical decision simulation, the model-assisted strategy increased surgical decision accuracy from 75% to 88%. 15 WSIs classified as PIL on FS but upgraded to IAC on permanent sections by pathologist were corrected by CryoPath and Attention heatmaps generated by the model highlighted histologically relevant regions. CryoPath model achieved an AUC of 0.92 (95% CI: 0.90–0.94) for discriminating IAC from PIL/MIA on the test set. . In clinical decision simulation, the model-assisted strategy increased surgical decision accuracy from 75% to 88. Attention heatmaps generated by the model highlighted histologically relevant regions. Conclusions: The CryoPath system demonstrates for the first time that large-scale deep learning-based analysis of FS-WSIs can achieve high-accuracy, real-time, and interpretable intraoperative diagnosis for LUAD. This system has the potential to serve as a powerful adjunct tool for pathologists and surgeons, potentially reducing diagnostic variability, shortening intraoperative wait times, and providing an objective foundation for precise surgical strategy. A prospective, multi-center clinical trial is being planned to further evaluate its clinical utility.