DIABETIC RETINOPATHY SEVERITY CLASSIFICATION USING CUSTOMIZED CNN

Abstract

Diabetic retinopathy (DR) is caused by diabetes, and could lead to permanent blindness. Diabetes causes damage to the arteries and veins of eye and further results in the loss of vision. Detection of DR is challenging as there are not many symptoms of the disease at the early stage. Opthalmologists can detect the DR, and analyze the severity by visual analysis of the fundus images. The population of diabetes-affected people is large and hence the manual detection and analysis is tedious and expensive. The automated screening and severity detection systems are therefore needed. In this paper, the experiments with CNN for classifying the DR images into five classes (NO DR, mild, moderate, severe and PDR) are conducted on the 2015 and 2019 Blindness Detection Kaggle dataset. Customized CNN model is designed in order to provide accurate severity classification. The CNN model consists of two convolutional layers, two max-pooling layers, one flattening layer and two dense layers. The retinal fundus images present structural and impulsive noise. Gaussian blur technique is applied as preprocessing to reduce the noise. We experimented with Adam’s optimizer as well as SGD optimizer and observed that the best classification results obtained using Adam’s optimizer (89% accuracy) are promising but warrant further investigation. Moreover, our approach has been able to have good and consistent AUC for all the classes, and particularly worked better than the existing approaches for the detection of class 3 (severe DR).