A Deep Learning Model for Diabetic Retinopathy Classification

An early diabetic retinopathy (DR) diagnosis leads to a reduction in the risk of blindness in diabetic disease patients. With the help of artificial intelligence, the efficiency is increased, and the cost is reduced for DR disease. One of the central ideas of deep learning is to take knowledge from a trained model on a dataset and apply it to a similar dataset with the same problem. This technique is called transfer learning. In this paper, we present a deep learning workflow that uses transfer learning to train an Xception model on our dataset with new customized layers. Transfer learning is used to reduce the required training data and minimize the learning time. We improved our model to obtain a high score of training accuracy (94%), test accuracy (0.89) and F1 Score (0.94) dealing with the unbalanced dataset of APTOS 2019 Blindness Detection.