Classificação de câncer de pele usando redes neurais convolucionais: uma análise do desempenho de classificação em um conjunto de dados desbalanceado

Abstract

Cancer is a leading cause of death worldwide, responsible for the death of approximately 9.8 million people just in 2018. Skin cancer is the most common type it’s desease, representing at least 40% of all cases of diagnosis. Given that the detection of early cancers is very relevant — cure rates reach 90% for low-risk melanomas — the use of techniques for rapid diagnosis is of great contribution to the area. The development of CNNs (Convolutional Neural Networks) that classify these pathologies is a resource that can be valuable for the development of auxiliary tools. The present work has implemented four CNNs: VGG16, ResNet-50, Resnet101 and Inception -Resnet, to classify skin cancer images from a HAM10000 database, in order to analyze and compare their performance in terms of precision, sensitivity and specificity, in five different scenarios. Was applied geometric transformations and transfer of learning. The network that obtained the best results was VGG16, both in relation to the general precision (obtained in the first scenario), with 82 %, and to the metrics analyzed for each class (fifth scenario) separately.