Reducing Overfitting in Diabetic Retinopathy Detection using Transfer Learning

Diabetic retinopathy is a diabetes complication which causes damage to the retina. It is highly prevalent in patients who have had diabetes for 20 years or more. Current deep learning methods of detection of this disease are not efficient with the respect to the number of images required for training and the transfer learning methods use pre-trained networks, which contains many features that are irrelevant to the original task data. In this paper, we pre-train our network in a self supervised manner on the training data, such that the network learns features relevant to the task data without overfitting. We propose a data efficient architecture based on autoencoder and VGG network and a method to train the network such that it is regularized by an autoencoder network to prevent overfitting and improve test data performance.