A Comparative Analysis of Glaucoma Feature Extraction and Classification Techniques in Fundus Images

Abstract

Glaucoma is a chronic asymptomatic eye disease that, if not treated in the initial stages, can induce blindness. However, early detection and proper treatment can prevent vision loss. Therefore, this work aims to evaluate the identification of glaucoma by non-invasive methods in fundus images. Initially, we have extracted the characteristics of images from the REFUGE and ACRIMA databases through the descriptors: Local Binary Patterns (LBP), Oriented Gradient Histogram (HOG), Zernike moments, and statistical information after the application of the Gabor filter. Then, we are given these characteristics to Multilayer Perceptron (MLP), Support Vector Machine (SVM), and Extreme Gradient Boosting (XGB) classifiers. The ranking is performed by each classifier individually and using a voting classifier. Additionally, we apply the cut-off threshold to define the predicted output due to the unbalance of the classes. To compare the results, we applied nonparametric tests. The voting classifier results reach an average rate for balanced accuracy equal to 93.29 \%, precision 88.74 \%, recall 92.04 \%, specificity 94.54 \%, and F2 score 91.33 \%. Therefore, using the cut-off threshold is essential for improving the recall results and reducing false negatives.