Diabetic Retinopathy classification through fundus images using Deep Learning

Abstract

One of the most common eye diseases in the people aged between 20-74 years is Diabetic Retinopathy (DR). DR is an eye complication where the patient loses his vision due to an increase in glucose levels in the blood. DR is most prominent in the patients who are diagnosed with the diabetes disease. Over one-third of the diabetic mellitus patients are diagnosed with DR. For diagnosing DR, the patient has to visit an ophthalmologist for dilated eye examination. However, everyone cannot have this facility. Hence, there is a need for a simple automated software for diagnosing the five stages of DR efficiently. In this paper, a simple model is developed using the Kaggle APTOS Blindness Detection dataset which is publicly available. In the pre-processing step the images are enhanced and the deep learning model ResNet152 architecture is used for the classification step. After training, the ReseNet152 model yielded a training and validation loss of 0.073 and 0.107 respectively and validation accuracy of 0.97. Further, a simple Graphical User Interface is developed using tkinter framework in python standard library which classifies the given input (a) (b) (c) fundus image as one of the five stages of DR.