Abstract:
Diabetic retinopathy (DR), a complication of diabetes, is one of the leading causes of blindness globally. Since early detection of DR can reduce the chance of vision loss significantly, regular retinal screening of diabetic patients is an essential prerequisite. However, due to inefficient manual detection as well as lack of resources and ophthalmologists, early detection of DR is severely hindered. Moreover, subtle differences among different severity levels and the presence of small anatomical components make the task of identification very challenging. The objective of this study is to develop a robust diagnostic system through integration of state-of-the-art deep learning techniques for automated DR severity detection. We used the concept of deep Convolutional Neural Networks (CNNs), which have revolutionized different branches of computer vision including medical imaging. Our deep network is trained on the largest publicly available Kaggle data set using our very own novel loss function. After several preprocessing and augmentation, 159,464 images were used for the training of the model. 10,000 images of Kaggle data was kept separate for testing purpose. Unlike most retrospective studies which perform binary classification (DR vs no DR), our model is trained to output five classes of DR severity as per international standard. An accuracy of 79.57% with a sensitivity of 79.58%, specificity of 82.81%, precision of 79.57% and F1 score of 0.778 was achieved on the test data. The model is also validated using two independent databases: Messidor and E-Ophtha to demonstrate its efficacy and generalization ability. In addition, a general comparison with some existing studies has been carried out to show that our model’s performance is comparable to the recent state-of-the-art models. The implementation of such a model to identify DR severity level accurately can reduce the risk of vision loss drastically by referring the affected to an ophthalmologist for further screening and treatment.
