An Ensemble Deep Learning Approach for Diabetic Retinopathy Detection using Fundus Image

Abstract

Detection of diseases, including diabetic retinopathy, may be greatly improved by taking a fundus picture of the back of the eye (DR). Complications in diabetics are the most common cause of vision problems, notably in younger and much more financially secure age groups. The risk of blindness in patients with DR may be reduced if they are diagnosed early enough. An ophthalmologist examined the fundus picture and used DR screening to look for lesions. However, the increase in incidence of DR is not correlated with the number of ophthalmologists who are able to interpret fundus pictures. Delay in prevention and treatment of DR may result as a result of this. Consequently, an automated diagnosis system is required to assist ophthalmologists in increasing the diagnostic process efficiency. The concatenate model is used in this study to differ fundus images into three categories: those without diabetic retinopathy, those with non-proliferative diabetic retinopathy, and those with proliferative diabetic retinopathy. We're using DenseNet121 and Inception-ResNetV2 for our models. Two models' feature extraction findings are integrated using the multilayer perceptron (MLP) classification approach. Compared to a single model, our strategy provides an increase in accuracy, precision, and recall of 91 percent and 90 percent for the F1-score. Deep-learning-based DR categorization utilizing fundus picture data was successfully shown in this experiment.