Towards Explainable Deep Learning for Non-melanoma Skin Cancer Diagnosis

Anh Le Van, Karin Verspoor, Thomas B. Kirk, Andy Song

Abstract: Skin cancer is a global health concern, but early diagnosis can be costly and challenging. An automated system is essential to prevent fatalities. This study focuses on non-melanoma skin cancer (NMSC), the most prevalent and one of the most serious health problems in the world, whereas the majority of computer-aided skin cancer diagnosis systems primarily target melanoma. Basal cell carcinoma (BCC), a kind of NMSC, is expected to overtake all other cancers in the near future. This study presents a highly accurate and automated diagnosis of NMSC. In addition to developing a classification model to differentiate NMSC from other skin diseases, we investigate the explainability of the model, since models should be accompanied by intuitive, persuasive, and consistent explanations for clinicians to accept the diagnosis. The model was trained and tested on the International Skin Imaging Collaboration (ISIC) archive and tested using the ISIC and derm7pt data sets, achieving high accuracy, sensitivity, and specificity. To deconstruct the prediction output, the model is studied using explainability methods that include gradient-based and perturbation-based approaches.