Integrating Non-local Information with Residual Blocks to Generate Explainable Attention Maps of Skin Cancer

Due to the rapid increase in population density, the field of medical sciences is currently confronted with a significant challenge regarding the automatic diagnosis of disease. In addition to assisting medical personnel in the early discovery of diseases, intelligent systems also contribute to the provision of consistent therapy, which in turn helps to minimize the death rate. In terms of mortality rate and severity, skin cancer is widely regarded as the most dangerous type of cancer. Images obtained through dermoscopy are utilized by medical practitioners to manually diagnose cases of skin cancer. This approach necessitates considerable skill, along with substantial time and labor dedicated to its implementation. When it comes to the early identification of skin cancer, automated detection technologies are absolutely required. An integration of non-local information with the residual attention block is provided in this article. The residual attention block is comprised of two sub-blocks, which are referred to as the channel attention block and the spatial attention block. It has been made possible for the proposed model to also provide attention maps because of the incorporation of residual attention with non-local information retrieval. Two classes, NV and Mel, from the ISIC2019 and Ph2 datasets are used to train the proposed model, and the ISIC2019 dataset is the sole one that is used for further evaluation. The experimental results indicate that the proposed approach attains superior performance and generalization in comparison to non-interpretable models, even when the method is not supervised by an expert.