Detection of basal cell carcinoma on whole-slide images from Mohs micrographic surgery using weakly supervised learning.

BACKGROUND: Mohs micrographic surgery (MMS) is the gold standard for treating aggressive basal cell carcinoma, but its success depends on expertise in intraoperative interpretation of frozen sections. OBJECTIVE: To develop weakly supervised, multiple instance learning framework using a pathology foundation model for automated basal cell carcinoma detection in MMS frozen sections. METHODS: An internal data set of 995 frozen MMS whole-slide images was slide-level labeled as tumor (512) or no tumor (483). Furthermore, tumor regions in the test set were annotated. Whole-slide images were tiled and encoded with Prov-Gigapath features for a weakly supervised multiple instance learning framework. The performance was evaluated as binary slide-level classification and in attention maps showing the localization of the tumor regions prior to validation on 2 external data sets. RESULTS: The model showed near-perfect diagnostic performance, achieving 97.0% accuracy, and an area under the receiver operating characteristic curve of 0.998 on the internal data set. Furthermore, attention maps visualized diagnostically relevant regions, enhancing model interpretability (Intersection-over-Union = 0.41 ± 0.046 [Dice 0.58 ± 0.046]). External validation confirmed robust performance (90% to 92% accuracy, area under the receiver operating characteristic curve: 0.92-0.95). LIMITATIONS: The absence of tumor region annotations on external data sets. CONCLUSION: These findings support the feasibility of artificial intelligence-assisted analysis of MMS frozen sections and justify prospective studies evaluating its integration into clinical workflows.