Automatic Discrimination between Normal and Diabetic Retinopathy Fundus Images using Machine Learning Classification Algorithm

Abstract

Diabetic retinopathy (DR) is a retinal disorder of the human eye caused by a complication of diabetes. It causes gradual damage to the eye’s retina, which can eventually lead to loss of vision. Early diagnosis of retinopathy is necessary for timely recovery of eyesight and to prevent complete blindness. In clinical routine, highly qualified specialists analyze the colored fundus images of each diabetic patient for the diagnosis of this disease. However, DR is not easily recognizable, particularly in the early stages, and even experienced experts may find such manual diagnostic procedures challenging, time-consuming, and error-prone. The goal of this study is to develop a robust automated system for classifying a given set of fundus images as normal or with DR symptoms, which would improve the performance of existing approaches. This work comprises three main stages: image preprocessing, feature extraction, and different supervised learning models for classification. A total of 65 features from each fundus image have been extracted and used as neural network (ANN) and support vector machine (SVM) inputs for the classification. Whereas high-level features are extracted from the ‘fully connected’ layer of a Residual Convolutional Neural Network (ResNet50) transfer learning model and utilized as the input feature vector for SVM classification. Based on the obtained findings, it can be indicated that the three approaches have successfully classified the images, but the latter has achieved the best classification accuracy of up to 98.38%. The results showed that these techniques for the classification of retinal fundoscopic images are fairly accurate.