Hyperparameters Optimization of Ensemble-based Methods for Retina Image Classification

Abstract

Diabetic retinopathy causes damage to the eye's retina and leads to visual impairment in diabetic patients worldwide. It affects the retina, begins asymptomatically and can lead to vision loss. It can be diagnosed quite accurately by using machine learning algorithms to analyze retina images. Diagnosis at an early stage is crucial to prevent dangerous consequences such as blindness. This paper presents a comparative analysis of ensemble machine learning algorithms and describes an approach to the selection of hyperparameters to solve the problem of diabetic retinopathy stage classification (from 0 to 4). Special attention is focused on grid search and random search approaches. This study proposed a hyperparameter selection technique for ensemble algorithms based on the combination of grid search and random search approaches. Hyperparameter selection increased retina image classification accuracy. Experimental results shown that hyperparameter selection increased retina image classification accuracy for testing dataset from 0.7460 for best model (GB) with default parameters to 0.7503 for best model (RF). If we consider binary classification (diabetic retinopathy presents or not) it is possible to achieve accuracy of about 0.9304 (RF).