Track

Abstract

Frozen section diagnosis of basal cell carcinoma (BCC) presents challenges for pathologists due to time constraints and the need for rapid, accurate intraoperative assessment in critical tissues of the head and neck. Objective: To develop and validate FROST, a deep learning-based classifier optimized for high sensitivity for automated detection of BCC in frozen section whole-slide images. Methods: A search was performed in our local database for frozen sections with a diagnosis of basal cell carcinoma. Positive cases and negative controls were identified within 403 individual slides from 41 cases and scanned using a portable digital slide scanner (Grundium Ocus®, Finland). WSIs were processed by extracting 224x224 pixel tiles at 20x magnification, and the Virchow 1.0 model was used to generate 2560-dimensional embeddings. These were used to train a 27.7 million-parameter classifier featuring a local transformer attention module that applies 8-head self-attention across 3x3 spatial grids of neighboring tile embeddings. The model was trained on 30 BCC-positive slides, 30 normal frozen sections, and 10 hard-negative regions. Evaluation was performed on 211 slides spanning a 6-year archival period. Results: FROST achieved an AUC of 0.978. At the optimized operating point, it demonstrated 100% sensitivity (58/58 positive cases) and 86.3% specificity (132/153 negative cases). Outputs are delivered as interactive HTML reports with tumor prediction overlays and adjustable thresholds for sensitivity-specificity trade-offs. Conclusions: We successfully developed and validated FROST, a screening tool optimized for high sensitivity. FROST serves as a safety net that works alongside the pathologist, boosting diagnostic confidence without compromising expert oversight.