NEURAL NETWORKS: DETECTING DIABETIC RETINOPATHY IN RETINA IMAGES

Abstract

The goal of this project is to develop an automated method to detect Diabetic Retinopathy (DR) in retina images using Convolutional Neural Networks. Diabetic Retinopathy is the leading cause of blindness in the working-age population of the developed world and it is estimated to affect over 93 million people. Designing a neural network that can diagnose and predict the occurrence of Diabetic Retinopa- thy in patients using retina images will speed up treatment and significantly im- prove the patients’ health condition. Currently this represents a laborious, manual task that can only be performed by qualified doctors, and the need for an automated solution has long been established. In this project, we build a Convolutional Neural Network using Tensorflow and TFLearn and train it on an Amazon Web Services GPU-enabled EC2 instance (p2.xlarge). We use the Diabetic Retinopathy Kaggle Competition whicg consists of 35.000 high resolution eye images for training and 55.000 images for testing. We experimented with a number of different Neural Net architectures including ConvNets, AlexNet, GoogleNet, Inception-v4, VGG-16 and ResNeXT. Our best results achieve a weighted-kappa score of 0.78601 which ranks 15 out of 661 participants, making it one of the top 2% best submissions on the Kaggle Leaderboard. In addition, we discuss potential commercial applications of these results as an iOS mobile application that allows users to take a picture of their retina using their iPhone camera, upload it to an AWS backend service, and find the probability of presenting signs of Diabetic Retinopathy.