Comprehensive Summary
The study presents XpressO-Melanoma, an interpretable deep learning (DL) model designed to predict BRAF V600E mutation status directly from hematoxylin and eosin (H&E)-stained whole slide images (WSIs) of cutaneous melanoma. Using 192 WSIs from the TCGA-SKCM cohort (112 BRAF wildtype and 80 BRAF-mutant cases), the authors trained a weakly supervised CLAM-based attention model within the XpressO pipeline to localize and interpret morphological features predictive of mutation status. On an independent test set (n=19), the model achieved an AUC of 0.8, precision of 0.72, and recall of 0.68. Four interpretive categories were established-true optimistic, false positive, true negative, and false negative predictions-each correlated with pathologist-reviewed histologic hallmarks. The attention heatmaps aligned strongly with tumor regions exhibiting high mitotic activity, eosinophilic nucleoli, and compact tumor nests, validating morphological plausibility in most correctly classified BRAF-mutant cases.
Outcomes and Implications
This study pioneers an interpretable deep learning framework for computational pathology, demonstrating that visual attention mechanisms can reveal biologically meaningful cues in melanoma histology. The model’s transparency bridges the gap between AI predictions and clinical interpretability, enabling pathologists to visualize the features that drive molecular inference. By achieving high accuracy using only H&E slides, XpressO-Melanoma offers a tissue-sparing and cost-efficient alternative to molecular assays such as RT-PCR or NGS, particularly in low-resource settings. Importantly, attention visualization identified cytologic correlates of BRAF mutation (e.g., pleomorphism, eosinophilic nucleoli, necrosis) and pinpointed cases in which AI focused on non-tumor regions, guiding future annotation refinement. These findings align with the White House National AI Action Plan's emphasis on interpretable and safe medical AI. Overall, this interpretable DL pipeline could accelerate precision oncology diagnostics, promoting human-AI collaboration in digit pathology while ensuring clinical trust and regulatory transparency.