Lightweight Convolution Neural Network for Diabetic Retinopathy Grading

Prevalence of Diabetic Retinopathy is very high in India. Regular screening is necessary for early detection and disease management. Screening programs generate large number of images and manual examination is time consuming and tedious. Traditional classification methods based on segmentation/handcrafted features methods present high false positive rate at pixel level. Deep learning models have been trained with massive datasets. The use of pretrained models on another task has limitations unless initial and target problem are similar enough. Medical images like digital fundus images have limited similarity with the imagenet database. In this study, authors propose to build and optimize lightweight CNN for grading of diabetic retinopathy into one of five classes. The model thus developed, is compared with pre-trained deep learning architecture Resnet. An accuracy of 83.51% is obtained with CNN and accuracy of 76.17% obtained with Resnet. Results demonstrate CNN can obtain better performance than pretrained models.