Enhancing Diabetic Retinopathy Detection Accuracy with Convolutional Neural Networks

Diabetic Retinopathy (DR) has emerged as a major illness that makes diabetic patients blind worldwide. If this disease is not treated promptly, it can progress to permanent blindness. Having the ability to identify people with DR rapidly will be beneficial for both medical professionals and researchers. In the current situation, manually analysing each minute nerve cell from fundus pictures is extremely time-consuming for clinicians. This paper not only concentrates on diabetic retinopathy detection but also on the analysis of DR progression in the given fundus image, which is performed with the help of data augmentation, Convolutional Neural network and transfer learning algorithms. Deep learning models DenseNet121, DenseNet169, DenseNet201, Resnet50, and InceptionV3 models are used on the Diabetic Retinopathy dataset in Kaggle and APTOS. Five DR stages, which are no DR, mild DR, moderate DR, severe DR and proliferate DR, are handled in this work. The fundus images of the eye are used as input to the model. The most robust model in the evaluations, DenseNet201, achieved an accuracy of approximately 98.56% for the APTOS dataset and 97.14% for the Diabetic Retinopathy dataset in Kaggle. These results were obtained using 5-fold stratified cross-validation and hyperparameter tuning, for detecting retinopathy and assessing its stage.