Classification of diabetic retinopathy through identification of diagnostic keywords

Abstract

Diabetic retinopathy is a condition caused due to diabetes affecting the blood vessels in the retina. This paper presents a two-phase approach for diagnosing various conditions of the eye and also classify the fundus image as diabetic retinopathy positive or normal. The ODIR dataset containing fundus images of various conditions is used for training and testing purposes. The proposed method consists of an ensemble model. The first phase is a convolutional neural network that takes fundus images for its input and outputs the diagnostic keywords for each eye. The second phase is a machine learning classifier that determines if a person has diabetic retinopathy or not based on the keywords generated from the previous model. The results of the two phases are satisfactory. The diagnosing phase has an accuracy up to 95% and the classifier has an accuracy up to 99%.