Técnicas de Aprendizado de Máquinas na Segmentação Automática dos Vasos Sanguíneos da Retina

Abstract

The images of retinographies, or images of the fundus of the eye, are obtained through the pupil and provide a direct representation of the nervous and vascular system of the human body, they are easily altered in the presence of different pathologies, hence the great interest in retinography images. Diseases such as lupus, diabetes, Zika, glaucoma, among others, produce significant changes in anatomical structures. In this sense, anatomical landmarks such as the optic disc, the macula and blood vessels are important markers in ophthalmologies exams. Consequently, many specialists apply digital image processing techniques in order to analyze, locate and identify such structures. In particular, the segmentation of blood vessels is one of the key elements to compose automatic screening or diagnosis systems. For example, the detection of hemorrhages present in the vascular structure of the retina can indicate a disease such as diabetic retinopathy, cardiovascular diseases, among others. Therefore, the general objective of this work is to develop and evaluate strategies based on Artificial Neural Network (ANN) and Convolutional Neural Network (CNN), in order to increase the performance in fundus image segmentation with an emphasis on blood vessels in terms of sensitivity, specificity and precision assessment metrics used in medical imaging. Two approaches are proposed to segment blood vessels in fundus images, so that the contributions can be listed as: (i) adaptation of a model Multilayer Perceptron (MLP) to detect the blood vessels, with results equivalent to the state of the art and (ii) adaptation of the CNN model, UNet architecture, for blood vessel segmentation, with the differential of using all the information of the RGB color model. The proposed methodology was evaluated in the publicly available databases: Digital Retinal Images for Vessel Extraction (DRIVE) and Structured Analysis of the Retina (STARE). The MLP and CNN UNet architectures provided expressive results in relation to the state of the art, with precision measurements of 0.9487 and 0.9713 in the DRIVE database.