High-Accuracy Binary Classification of Diabetic Retinopathy from Fundus Images Using CNN Architecture

Abstract

Diabetic Retinopathy (DR), a retinal condition associated with diabetes mellitus, is a leading global cause of blindness. Early detection and treatment is crucial to prevent or mitigate vision loss. Researchers have developed various artificial intelligence-based methods to detect diabetic retinopathy from fundus images, aiming to enhance diagnostic accuracy and efficiency, thereby improving patient outcomes. In this paper, we have introduced a custom convolutional neural network (CNN) architecture for detecting diabetic retinopathy. We have used an open-source “Diabetic Retinopathy 224×224 Gaussian Filtered” dataset from Kaggle, comprising of 3,668 images. The images are then labelled into two classes where one class represents “No Diabetic Retinopathy (No_DR)” and the other class represents “Diabetic Retinopathy (DR)”. Then we have applied min-max normalization to scale pixel values to a specified range [0, 1]. Afterwards, the dataset is divided into three parts to train, validate and test the proposed CNN model. Our method has achieved a testing accuracy of 95.27% and F1 score of 95.29%. Then we have performed 5-fold cross validation to observe the performance of the proposed automatic DR detection method. We have also compared our proposed method with some pre-trained models and some existing methods. In observation, it has been shown that the proposed technique outperformed many other existing techniques.