Diabetic Retinopathy Fundus Image Classification Using Ensemble Methods

Abstract

Diabetic retinopathy causes damage to the retina of the eye and leads to poor vision in patients with diabetes around the world. It affects the retina of a person’s eye, begins asymptomatically, and can lead to complete loss of vision. Screening for this disease can be done fairly quickly by using machine learning algorithms to analyze retinal images. Early diagnosis is crucial to prevent dangerous consequences such as blindness. This paper presents the results of implementation and comparison of ensemble machine learning algorithms and describes an approach to the selection of hyperparameters for solving screening problems (binary classification) and classifying the stage of diabetic retinopathy (from 0 to 4). Particular attention is paid to the approaches of searching for hyperparameters on a lattice and random search. This study uses a hyperparameter selection mechanism for ensemble algorithms based on a combination of grid search and random search approaches. The selection of hyperparameters, as well as the selection of informative features, made it possible to increase the accuracy of classification of retinal images. The experimental results showed an accuracy of 0.7531 for retinal image classification on the test dataset for the best model (gradient boosting, GB). When considering a binary classification (presence or absence of diabetic retinopathy), an accuracy of 0.9400 (gradient boosting, GB) was achieved.