Uma abordagem de classificação de câncer de pele usando GAN e mecanismo de atenção baseado em RoI

Abstract

Skin cancer is a global health problem, being the most common type of cancer, being divided into two categories, non-melanoma cancer, more common and less lethal, and melanoma cancer, less common, but with a high mortality rate. Early diagnosis is the best way to fight skin cancer, avoiding invasive methods in non-melanoma skin cancer and increasing the cure rate and survival in the case of melanoma type cancer. Several computer vision techniques are being used in the medical field, in order to help professionals with diagnosis, treatment recommendation, among others. One of these methods is the use of convolutional neural networks, also called by the acronym CNN, for the classification of images of lesions and tumors, with studies showing an even greater accuracy capacity than trained physicians. However, some problems are encountered when it comes to skin lesion image classification, such as image set imbalance. For this, several techniques can be used for the generation of images in order to balance the set, however, one that has gained prominence is the use of generative adversarial networks, also called by the acronym GAN, which are capable of generating synthetic images with high quality based on a set previously used in a training process. This work aims to search for methods that lead to an increase in the performance of CNN networks for the classification of skin lesions. For that, this work proposes a CNN architecture based on the EfficientNetB0 network, called EfficientAttentionNet, for classification of skin lesions, specifically melanoma and non-melanoma. First, the original image dataset, from the International Society for Digital Skin Imaging (ISDIS), is pre-processed to eliminate hairs around the skin lesion. Subsequently, a GAN model generated synthetic images to balance the number of samples per class in the training set. A U-net template is used to create masks with the region of interest of the image. Finally, the proposed EfficientAttentionNet model to classify skin cancer using mask attention mechanisms is presented. The results showed that the proposed classification model achieved high performance results, obtaining accuracy of 0.979, precision of 0.945, recall of 0.995 and ROCAUC of 0.976, serving as a reference for research in the area of classification of skin lesions.