Attention-guided deep convolutional neural networks for skin cancer classification

Skin cancer is a silently killing disease which commonly goes unnoticed in its primitive stage but proves to be deadly later on. Hence, it needs to be detected and classified in the early stages itself. The advent of machine learning as well as deep learning based classification techniques has made this task possible. Deep convolutional neural networks (D-CNNs) have the ability to extract universal and dataset-specific features for the image classification task. But the classification of skin cancer images remains a challenging task due to the absence of balanced class images, difference between images of the same class, similarity between inter-class images and the inefficiency in focusing on the semantically significant areas of the image. To improve the performance of these D-CNNs, we incorporate the attention mechanism that focuses on the regions of importance in an image. In that regard, we propose an attention-guided D-CNN for classification of skin cancer. It is observed from the classification results that a model with attention boosts the accuracy of a normal D-CNN architecture by approximately 12%. Our research work contributes significantly to the field of biomedical image processing by providing a mechanism to improve performance of D-CNNs and facilitating early detection of skin cancer.