CNN Based Detection of the Severity of Diabetic Retinopathy from the Fundus Photography using EfficientNet-B5

Diabetic Retinopathy (DR) is prevalent among a large number of population who has diabetes. It is a very critical disease which causes injury in the tissues of retina of diabetic patients. Even, at its extreme point, it might cause permanent blindness to the patients who suffer from DR for a long period of time. Therefore, it is necessary to diagnose these patients very soon to mitigate the severe impact of DR. Several methods are proposed earlier to identify this disease using machine learning algorithms, image processing, and so on. In this work, a Convolutional Neural Network (CNN) based pre-trained machine learning algorithm is used to speed up the diagnosis of the severity of DR using fundus photography of retinal images of patients. The images are of five different classes which defines the severity of the disease. We used EfficientNet-B5 along with an optimizing threshold to further improve the result. We achieved an accuracy of 0.9402 on the training set and 0.9333 on the testing set which is represented by Quadratic Weighted Kappa score. This proves the efficacy of our approach in DR classification. Thus, our work can help the DR patients to reduce the probability of having lifelong blindness.