Melanoma Skin Cancer Classification with Explainability

Abstract

Melanoma is a fatal skin cancer with a high prevalence worldwide. The likelihood that a melanoma patient would recover considerably increases with early detection. At present deep learning (DL) approaches are becoming popular in assisting early melanoma identification. Although DL techniques provide high performance, the utilization of an image classifier alone results in the low trustworthiness of the application and makes it challenging to grasp the reasoning behind model predictions. This emphasizes the requirement of justifying the decision in addition to the classifications with better performance. In contrast to existing black-box methods, this paper addresses the explainability of a classification. We present a computational model to classify melanoma skin cancer images by applying the Xception model for the HAM10000 dataset. With the retaining of batch-normalization layers and Bayesian hyperparameter search to fine-tune hyperparameters, this study shows a classification accuracy of 90.24%. Additionally, we generate heatmaps using Gradient-weighted Class Activation Mapping (Grad-CAM), and Grad-CAM++, for the explainability of the classification model, as a novel contribution to the domain of dermatology. The visualized heatmaps explain the contribution of each input region to the classification result.