Early-stage diabetic retinopathy severity detection using Generative Adversarial Network

Abstract

Diabetes is the most common cause of retinopathy which causes vision impairment. Available estimate on diabetic retinopathy shows that 6 million diabetes patients in India have a sight-threatening form of retinopathy. The state-of-the-art high performance Deep learning(DNN) architecture has been used to detect blood vessel abnormalities in the retinal tissue due to diabetes mellitus obtained from fundus camera. But the deep neural network suffers due to lack of generalization,  data imbalance, and difficulty in reaching the global optimum because of implication of less data. Larger amount of data impacts the performance of the model with meaningful feature extraction capability and not just noise so the model can learn from it. Synthetic medical data generation in medical imaging is still a challenging, expensive and important goal, which wanted to look like a real image. In this paper, to enhance the performance of the model and to overcome the dependencies of the less number of retinal images for training, Generative adversarial network (GAN) has been used to generate high resolution synthetic retinal images, where even the retinal specialist were unable to differentiate the synthetically generated image from the real retinal images. This will help the neural network model reach global optimum and aid in computer assisted medical image screening to improve the diagnostic reliability for clinical evaluation in the near technologically advanced healthcare.