Avaliação de métodos de classificação de glaucoma em imagens de fundoscopia

Abstract

Glaucoma is an asymptomatic eye disease that, if not treated on time, can lead to blindness. The World Health Organization (WHO) estimates that by 2020 glaucoma should affect 80 million people and by 2040 it will be 111.5 million. In this context, the present dissertation aims to compare classification methods and study different techniques for the extraction of image characteristics, thus assisting the specialist physician in diagnosing the disease. Three models are developed based on different types of feature extraction. Model 1 extracts nongeometric characteristics: Local Binary Pattern (LBP), Histogram of Oriented Gradients (HOG), Zernike moments and Gabor filter statistical information. Model 2 is similar to model 1 with the addition of geometric features. And in model 3, pre-trained convolutional network models (MobileNet, VGG16, VGG19 and Resnet50) are used to extract information from the images. For each model, the obtained characteristics are submitted to Principal Component Analysis (PCA) for dimensionality reduction, the resulting components are classified by: Logistic Regression (RL), Gradient Increasing Decision Tree (GBDT), Support Vector Machine (SVM), k-nearest neighbors (k-NN), and Multilayer Perceptron (MLP). To improve classification performance, hyperparameter optimization techniques using Grid Search are used. Of the three models evaluated, model 1 produces the best results using SVM for classification. The test results achieved an average accuracy rate of 89.03%, sensitivity of 86.59%, specificity of 91.06% and AUC (area under a curve) of 88.95%.