Early Detection and Severity Classification of Diabetic Retinopathy Using Convolutional Neural Networks

Abstract Diabetic retinopathy (DR) has become a leading cause of blindness, and detection of the condition at an early stage is important for successful treatment. Nonetheless, it is quite difficult to detect DR in its initial stages in areas with a lack of medical care. This research seeks to develop a neural network that will have the ability to (1) detecting the presence or absence of DR, (2) early, detection (3) classification of severity of DR. We used the APTOS DR dataset that contains 3681 fundus images with DR ratings from 0 (no DR) to 4 (severe proliferative DR). Three distinct models were trained: a binary classifier, an early detector, and a severity classifier that use a neural network with three convolutional layers, a global average pooling layer, and three fully connected layers. The models were cross-validated, with a fivefold used, tracking the training and validation accuracy. The binary classifier was able to have a validation accuracy of 96.2% and an AUC of 0.992, which is higher than existing models in the literature. Early detector managed to have 86% accuracy but had difficulty distinguishing between early and severe DR. The accuracy of the severity classifier was 79.4%, being very successful in detecting healthy subjects but failing to classify more severe cases, possibly because of the model’s inability to discriminate against slight differences between later DR degrees. Such results show the effectiveness of the NN usage in the diagnostics of DR and its classification, but still, more work is required for better severity prediction.