Prediction of BAP1 mutations in uveal melanoma patients from histology images using weakly supervised deep learning-based whole slide image analysis

AbstractWhile cases of uveal melanoma are relatively rare overall, it remains the most common intraocular cancer in adults and has a 10-year fatality rate of approximately 50% in metastatic patients with no effective treatment options. Mutations in BAP1, a tumor suppressor gene, have been previously found to be associated with the onset of metastasis in uveal melanoma patients. In this study, I utilize a weakly supervised deep learning-based pipeline in order to analyze whole slide images (WSIs) of uveal melanoma patients in conjunction with slide-level labels regarding the presence of BAP1 mutations. I demonstrate that there is a strong relationship between BAP1 mutations and physical tumor development in uveal melanoma and that my model is able to predict such relationships with an optimized mean test AUC of 0.86. My findings demonstrate that deep learning models are able to accurately predict patient-specific genotypic characteristics in uveal melanoma. Once integrated into and adapted to existing non-invasive ocular scanner technologies, my model would assist healthcare professionals in understanding the specific genetic profiles of their patients and provide more personalized treatment options in a safe, efficient manner, thus resulting in improved treatment outcomes.