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

Deep learning (DL) models have shown promise in predicting molecular alterations directly from hematoxylin and eosin-stained whole slide images in a variety of solid tumors, offering a rapid alternative to conventional molecular testing. However, these models often offer limited insight into their decision-making process, undermining transparency and eroding clinical trust. Interpreting model predictions is essential for a meaningful application of DL in clinical pathology. This is showcased by interpreting the outputs of a weakly supervised DL model, XpressO-melanoma, that predicts BRAF V600E mutation status from whole slide images of skin cutaneous melanoma. The morphologic plausibility of the model's segmentations of the tumor regions of interest and their prediction of BRAF V600E status were evaluated and compared against the pathologists' annotations for the same. The work resulted into four interpretation categories that associate model's performance (ie, area under the curve of 0.8 and precision and recall of 0.7) with the regions of interest that revealed meaningful diagnostic patterns as well as those that required annotation refinements. The work coheres with the White House's National AI [Artificial Intelligence] Action Plan that identifies interpretability as a national research priority and paves the way for a human-DL collaboration in clinical pathology for a better translation of DL techniques in clinical pathology in the near future.