A skin cancer classification approach using GAN and ROI-based attention mechanism

Abstract

Skin cancer is a complex public health problem and one of the most common types of cancer worldwide. A biopsy of the skin lesion gives the definitive diagnosis of skin cancer. However, before the definitive diagnosis, specialists observe some symptoms that justify the request for a biopsy and consider a early diagnosis. Early diagnosis of skin cancer is subject to errors due to the lack of experience of specialists and similar characteristics with other diseases. This work proposes a CNN architecture, called EfficientAttentionNet, to provide early diagnosis of melanoma and non-melanoma skin lesions. The methodology represents the stages of development of the proposed classification model and the benefits of each stage. In the first step, the set of images from the International Society for Digital Skin Imaging (ISDIS) is pre-processed to eliminate the hair around the skin lesion. Then, a Generative Adversarial Networks (GAN) model generates synthetic images to balance the number of samples per class in the training set. In addition, a U-net model creates masks for regions of interest in the images. Finally, EfficientAttentionNet training with the mask-based attention mechanism to classify skin lesions. The proposed model achieved high performance, being a reference for future research in the classification of skin lesions.