Identificação do glaucoma em imagens do fundo do olho utilizando aprendizagem profunda

Abstract

Glaucoma is a disease that damages the optic nerve and can cause loss of vision or total blindness. This disease is the major cause of irreversible blindness in the world. It is estimated that by 2020 the number of people with glaucoma could reach 76 million. This work compares the accuracy of different neural network architectures that use deep learning for image recognition. These neural networks can help healthcare professionals to diagnose glaucoma more efficiently and precisely, since the process is done manually by specialists. This work utilizes a state-of-the-art, high-performance object detection system called the YOLO9000, responsible for detecting the optic nerve, which is the region of interest. After detection of this region, a convolutional neural network was used to detect the presence of glaucoma. This work analyzes different classifiers to verify which one has the best accuracy to solve this problem. Public available fundus images databases were used to validate this process. The convolutional neural network called DenseNet was the one with best average accuracy to detect the glaucoma among the used images databases. The results were 100 %, 85.8 94.4 % in the HRF, RIM-ONE-R1-R2 and RIM-ONE-R3 image databases, respectively, using the area under the receiver operating characteristic metric.