Classification of Diabetic Retinopathy and Retinal Vein Occlusion in Human Eye Fundus Images by Transfer Learning

Abstract

Sight threatening diseases are viral these days. Some of these are so harmful that it may cause complete vision loss. Diabetic Retinopathy (DR) and Retinal Vein Occlusion (RVO) are from this category. The first step to cure such diseases is to accurately predict it. For prediction of these diseases there is a large number of machine/deep learning algorithms employed. In this research, we have proposed a DR&RVO prediction system, which may help eye specialists for the prediction of these diseases. The proposed methodology shows that a retinal image undergoes through three main steps in a Deep Neural Network (DNN) like, preprocessing, image segmentation, and feature extraction and classification. For classification of this processed image into DR and RVO, and normal labels, pre-trained deep neural networks (DNNs) are used. More than 2680 eye fundus images are collected from 7 online available datasets, all images are converted to jpg file format during preprocessing step, after class labels distribution into three categories, the proposed model is firstly trained and then tested randomly on Inception v3, ResNet50 and Alex Net. This is done by using a deep learning technique named as ‘Transfer Learning’. The accuracy obtained from these models shows that Inception v3 (85.2%) outperformed than other two state of the art models.