An XAI Approach to Melanoma Diagnosis: Explaining the Output of Convolutional Neural Networks with Feature Injection

Abstract

Computer-aided diagnosis (CAD) systems, which combine medical image processing with artificial intelligence (AI) to support experts in diagnosing various diseases, emerged from the need to solve some of the problems associated with medical diagnosis, such as long timelines and operator-related variability. The most explored medical application is cancer detection, for which several CAD systems have been proposed. Among them, deep neural network (DNN)-based systems for skin cancer diagnosis have demonstrated comparable or superior performance to that of experienced dermatologists. However, the lack of transparency in the decision-making process of such approaches makes them “black boxes” and, therefore, not directly incorporable into clinical practice. Trying to explain and interpret the reasons for DNNs’ decisions can be performed by the emerging explainable AI (XAI) techniques. XAI has been successfully applied to DNNs for skin lesion image classification but never when additional information is incorporated during network training. This field is still unexplored; thus, in this paper, we aim to provide a method to explain, qualitatively and quantitatively, a convolutional neural network model with feature injection for melanoma diagnosis. The gradient-weighted class activation mapping and layer-wise relevance propagation methods were used to generate heat maps, highlighting the image regions and pixels that contributed most to the final prediction. In contrast, the Shapley additive explanations method was used to perform a feature importance analysis on the additional handcrafted information. To successfully integrate DNNs into the clinical and diagnostic workflow, ensuring their maximum reliability and transparency in whatever variant they are used is necessary.