An evaluation of pathology foundation models with weakly supervised learning for BRAF mutation prediction in melanoma whole slide images

Abstract

Introduction Melanoma is the most aggressive form of skin cancer, with around 40–50% of cases harbouring BRAF mutations that can be effectively targeted to improve patient outcomes. Typically, BRAF mutations are identified through DNA sequencing, which can be costly and contribute significantly to turn-around-times. Predicting BRAF status directly from haematoxylin and eosin (H&E) whole-slide images (WSIs) could offer a faster and more accessible alternative within digital pathology workflows. Methods In this study, we evaluated six pathology foundation models for WSI-based BRAF mutation prediction using an attention-based multiple instance learning (ABMIL) framework. Training and testing models with five-fold cross-validation, using 1,109 WSIs from 745 patients and performing external validation with TCGA and VisioMel cohorts. Results Models which utilised H-Optimus-1 features achieved the highest AUROC values on the internal test set (0.80 (0.72, 0.88)) and VisioMel (0.79 (0.75, 0.83) external test set, while UNI2-h features demonstrated superior sensitivity and higher TCGA performance (AUROC: 0.76 (0.66, 0.86)). Subgroup analyses indicated that model performance was impacted by histopathological features such as ulceration and tissue area, while t-SNE plots showed that high-attention foundation model features tended to cluster based on the dataset the features were from, rather than biological signal. Discussion Overall, foundation models combined with weakly supervised learning provide competitive performance for BRAF status prediction in melanoma WSIs, but our findings highlight the need for diverse melanoma datasets and the importance of evaluating model performance beyond a single metric when developing H&E-based computational biomarkers.