An interpretable hierarchical semantic convolutional neural network to diagnose melanoma in skin lesions

Abstract

Melanoma is a more dangerous skin cancer than other types of skin cancer because it rapidly spreads to other organs in its early stages. In the increasingly popular task of computer-aided diagnosis using deep learning methods, these models are difficult to interpret and often considered "black boxes". The lack of interpretation of the model prevents the target users from fully understanding it. This study proposes a new interpretable hierarchical semantic convolutional neural network (MEL-HSNet) to diagnose melanoma. The benefits and strength of our approach are a white-box model that not only predicts whether a skin lesion observed in a dermoscopy scan image is melanoma but also provides explanatory information for decision-making. Compared to other convolutional neural networks, the MEL-HSNet model proposed in this study can generate interpretable information on melanoma prediction and obtain significantly better results compared to the other available models.

Keywords

• melanoma classification
• skin cancer
• deep learning
• convolutional neural networks
• model interpretability
• intelligent
• healthcare