Interpreting Deep Learning-Based Prediction of the BRAF V600E Mutation Using Diagnostic Whole Slide Images in Skin Cutaneous Melanoma

Comprehensive Summary

The following study investigates whether deep learning (DL) can predict BRAF V600E mutation status directly from hematoxylin and eosin (H&E)-stained whole-slide images (WSIs) of skin cutaneous melanoma (SKCM) while also making the model’s decision-making process interpretable. The study introduces XpressO-melanoma, a weakly supervised DL pipeline that segments tumor regions of interest and analyzes histologic patterns associated with BRAF mutations. Using 192 melanoma WSIs from The Cancer Genome Atlas (TCGA), the authors trained and tested the model to classify tumors as either BRAF V600E mutant (BVE) or BRAF wild type (BVW). The model achieved an AUC of 0.8 with precision and recall around 0.7, demonstrating moderate predictive performance. The dataset includes 154 training slides and 19 each for validation and testing. Attention heat maps were then generated to identify which tumor regions most influenced the predictions, allowing pathologists to compare AI-highlighted areas with manually annotated tumor features. This interpretability approach revealed that the model frequently focused on histologically meaningful tumor structures associated with mutation status.

Outcomes and Implications

The following study demonstrates that interpretable AI models can potentially predict clinically actionable genetic mutations directly from routine pathology slides, offering a faster and less resource-intensive alternative to molecular testing. BRAF mutations occur in approximately 40-60% of melanomas and guide treatment decisions involving targeted therapies such as BRAF and MEK inhibitors, making accurate identification clinically essential. The model correctly classifies 87.5% of BRAF-mutant tumors and 54% of wild-type tumors on a test set, indicating that, while promising, performance still varies with histologic complexity. The attention maps revealed that the algorithm recognized known morphological hallmarks of BRAF-mutant melanoma—such as prominent nucleoli, high nuclear-to-cytoplasmic ratios, and dense tumor cellularity—as shown in the example cases. Misclassifications often occurred when tumors exhibited overlapping morphologic features or when the model focused on non-tumor regions such as necrosis or stroma. The authors conclude that combining deep learning with pathologist-guided interpretation could improve diagnostic workflows, especially in settings where molecular testing is expensive or unavailable. Ultimately, interpretable AI systems like XpressO-melanoma may serve as decision-support tools to prioritize patients for confirmatory genetic testing and targeted melanoma therapies.