Detección de melanomas de piel malignos mediante procesamiento digital de imágenes usando redes neuronales convolucionales

Abstract

This project aims to design a neural network for the recognition of melanomas (a type of skin cancer), through the use of a technique known as convolutional neural networks, mostly used in artificial vision (a branch of artificial intelligence), applied in the recognition of patterns on moles on the skin and determine the existence of a malignant melanoma, or not, from a limited dataset. For this, the convolutional network designed and trained to classify melanomas is made up of convolution and pooling layers stacked together to form the proposed network, a "fully connected layer" and a classifier with 1 or 2 outputs, and is parameterized with different values ​​in characteristics such as the dropout, the size of the filters, among others, performing the training in 5 different stages or experiments. The dataset proposed for the training of CNN (Convolutional Neural Networks) is the largest public collection of demoscopic images of skin lesions, provided free of charge by the "International Skin Imaging Collaboration (ISIC)", an effort to improve the diagnosis of melanomas, sponsored by the “International Society for Digital Imaging of the Skin (ISDIS)”. The purpose of this project is to design a convolutional neural network with a high level of precision that helps medical professionals with the diagnosis of melanomas, in this case it was possible to achieve an accuracy of up to 87.82% with the network designed with the best performance.