Grading of Retinal Images Using CNN and its Variants

People having Type I or Type II diabetes may develop an eye syndrome called Diabetic Retinopathy. High blood glucose levels harm the blood vessels in the retina. Initially, it may be asymptomatic or may only result in minor eye problems. But, if left untreated may even steal the vision. Diabetic Retinopathy traditional diagnosis requires a great deal of time and effort from ophthalmologists. In this work, convolutional neural networks are used to automatically grade various forms of diabetic retinal disease. To improve the classification accuracy, various pretrained deep learning architectures like VGG-16 and six variants of ResNet namely ResNet 50, ResNet 101, ResNet 152, ResNet 50v2, Resnet101v2, ResNet152v2 are used in this study. Regularization technique called Earlystopping and hyperparameter tuning is also performed. Based on the conducted experiments, it has been identified that ResNet152V2 performs better than other proposed models, with a maximum accuracy of 98.7 when batch size is set to 32, optimizer RMSprop is used, and dropout rate is set to 0.000001.