Segmentação de vasos sanguíneos utilizando Redes Neurais Convolucionais: visualização e análise de correlação dos mapas de ativação

Abstract

Convolutional neural networks (CNN) are responsible for advancing the automation of tasks in the field of computer vision, with image classification, object recognition and image segmentation, as it is present in different research and industry areas. In particular, the biomedical area, which makes use of CNNs for segmentation of x-ray images and microscopy of cells and blood vessels. These networks are known to be deep, composed of several layers that extract features, called activation maps, from the images and generate complex models with millions of parameters. Therefore, it is important to develop approaches for visualization and interpretation of the generated activation maps. Furthermore, investigating the degree of similarity between representations can lead to a better understanding of how networks generate their information representations. This study investigates ways to extract representations produced by CNNs and develop methods and tools to visualize activation maps, in addition to calculating and analyzing the correlation of these maps in segmentation networks with different depths. Understanding the degree of correlation among the maps and identifying highly correlated maps can be useful for applying methods that make networks more computationally efficient without decreasing their accuracy. This work also proposes to investigate the degree of correlation and similarity between different maps produced by networks, either by a single network or by comparing networks of different depths. The study carried out found that CNNs produce redundant maps as they get deeper, tending to form blocks of highly correlated maps in the last layers.