Investigating on data augmentation and generative adversarial networks (GAN s) for diabetic retinopathy

The early detection of diabetic retinopathy (DR) disease, which is a complication of diabetes, may help to reduce blindness. One of the issues faced currently is the limited number of ophthalmologists available to take care of the diabetes patients. To overcome this problem, automated DR grading applications are required. Many researchers are now developing deep learning to build these decision support systems. Nevertheless, the frequent misclassification of DR remains an open challenge. Class imbalance in the datasets and limited labelled datasets are the root causes of misclassification. In this work, we are investigating data augmentation and Generative Adversarial Networks (GANs), known to countermeasure class imbalance and the dearth of labelled images. We have applied these two techniques on the APTOS dataset after the undersampling process. Traditional data augmentation has achieved an accuracy of 72 % versus GAN which has reached an accuracy of 76%.