An Ensemble Framework for Glaucoma Classification Using Fundus Images

Abstract

Glaucoma is an irreversible progressive vision condition that can lead to permanent sightlessness. With null early-stage symptoms, it is critical to prevent in advanced stages of glaucoma. Artificial intelligence has shown significant escalation, in many fields, especially in medical image diagnosis, wherein highly accurate automated disease diagnosis of a large dataset in less time has now become feasible. Researchers have utilized many machine learning and deep learning approaches for glaucoma detection from retinal fundus images, but their performance varies and it depends on the input dataset. According to the “No Free Lunch theorem”, a single classifier is not suitable for classifying all datasets. So, to overcome this situation, we have proposed an ensemble learning approach that utilizes different machine learning approaches for Glaucoma detection. A benefit of using ensemble learning is to improve the average prediction performance by combining predictions from multiple models. In this paper, we developed a machine learning model ensemble approach [Majority-Voting-Ensemble (MVE)] consisting of a Support Vector Machine (SVM), Random Forest (RF), Multilayer Perceptron (MLP) classifiers. Our approach involves the Histogram of Oriented Gradients (HOG) technique to collect efficient features for glaucoma detection. The efficiency of our proposed approach is evaluated using two popularly used benchmark datasets, called ORIGA and REFUGE. The results show that the proposed ensemble approach is capable of outperforming the base classifiers for glaucoma classification from fundus images.