Improved Diagnosis of Non-Melanoma Skin Cancer in Resource-Limited Settings.

Background: Early and precise diagnosis is vital to improving patient outcomes and reducing morbidity. In resource-limited settings, cancer diagnosis is often challenging due to shortages of expert pathologists. We assess the effectiveness of general-purpose pathology foundation models (FMs) for the diagnosis and annotation of nonmelanoma skin cancer (NMSC) in resource-limited settings.

Methods: We evaluated three pathology FMs (UNI, PRISM, and Prov-GigaPath) using de-identified NMSC histology images from the Bangladesh Vitamin E and Selenium Trial to predict cancer subtype based on zero-shot whole slide embeddings. In addition, we evaluated tile aggregation methods and machine learning models for prediction. Lastly, we employed few-shot learning of PRISM tile embeddings to perform whole slide annotation.

Results: We found that the best model used PRISM's aggregated tile embeddings to train a multi-layer perceptron model (MLP) to predict NMSC subtype (mean AUROC=0.925; p<0.001). Within the other FMs, we found that using attention-based multi-instance learning to aggregate tile embeddings to train an MLP model was optimal (UNI: mean AUROC=0.913; p<0.001; Prov-GigaPath: mean AUROC=0.908, p<0.001). We finally exemplify the utility of few-shot annotation in computation- and expertise-limited settings.

Conclusions: Our study highlights the important role FMs may play in confronting public health challenges and exhibits a real-world potential for machine learning aided cancer diagnosis.

Impact: Pathology foundation models offer a promising pathway to improve early and precise NMSC diagnosis, especially in resource-limited environments. These tools could also facilitate patient stratification and recruitment for prospective clinical trials aimed at improving NMSC management.