Diabetic Retinopathy (DR) is one of the leading causes of preventable blindness in the working-age diabetic population in India and across the world. It may lead to permanent blindness if not detected in the early stages. The prevalence of DR among diabetics in India was 10% and 16.9% in 2014 and 2019, respectively. In 2019, the International Diabetes Federation estimated that Diabetic Mellitus will affect 101 million people in India in 2030; the largest number in any nation in the world. Our work is an attempt to speed up preliminary screening of DR to cater to the future requirement of such a huge amount of diabetic patients. We have trained and validated robust classification models on publicly available datasets for early detection of DR. We have applied state-of-the-art deep learning models based on Convolutional Neural Networks (CNN), to exploit data-driven machine learning methods for the purpose. We framed the problem as a binary classification for the detection of DR of any grade (Grade 1–4) vs No-DR (Grade 0). We used 56,839 fundus images from the EyePACS dataset for training the models. The models were tested on a test set from EyePACS (14,210 images), benchmark test datasets Messidor-2 (1748 images) and Messidor-1 (1200 images). The model has achieved an AUC of 0.92 on benchmark test dataset Messidor-2 with sensitivity and specificity of 81.02% and 86.09%, respectively. AUC, Sensitivity and Specificity on Messidor-1 are 0.958, 88.84% and 89.92%, respectively. In this paper, we also discuss challenges of automated ailment detection in medical images using CNNs, such as the use of public datasets for training, pre-processing methods, performance metrics for unbalanced classes and present our results and their comparison with leading studies. The developed preliminary automated screening system will act as an aid to the manual diagnostic process by referring DR patients to an ophthalmologist for further examination (if detected positive) well in time to reduce the risks of vision loss.
Intelligence-Based Medicine, Volumes 3–4, December 2020, 100022,