Diabetic Retinopathy Detection from Fundus Images Using Deep Convolutional Neural Networks

Diabetic Retinopathy (DR) is one of the primary causes of blindness. The earlier it gets detected; the earlier patients get the treatment. Currently, diagnosis of DR depends on traditional manual methods and resources, which often leads to human error and wrong diagnosis. This issue can be avoided through the development of an automated system for accurately detecting DR. In this research work, we are proposing an approach to detecting DR through deep Convolutional Neural Networks (CNNs) trained using fundus images (photos of the retina of the eye with a fundus camera). The developed CNN model detected the severity of diabetic retinopathy from fundus images with the highest probability. The final model with the best performance was achieved through iterative fine-tuning of the hyperparameters and changing of the layers. The model was tested to avoid overfitting or underfitting issues. We developed models for supervised multiclass classification as well as binary classification, where ‘No DR’ was considered as a single class, and the other types of DR were combined to make a ‘DR’ class. In binary classification, the best model achieved a validation accuracy of 96% and testing accuracy of 95%. This study highlights the potential of deep learning models to automate the detection of diabetic retinopathy, offering the possibility of more efficient and scalable screening methods which could significantly reduce the incidence of blindness from this condition.