Diabetic retinopathy severity classification using convolutional neural network withtransfer learning

Abstract

Diabetic Retinopathy is a common retina disease caused by diabetes that is very difficult to diagnose initially because of its asymptomatic nature, which leads to permanent vision loss. Early and accurate detection of diabetic retinopathy is an effective way to prevent blindness. While screening programs are the most effective means of detecting diabetic retinopathy, it has various limitations like time consuming, laborious tasks, require a lot of expert ophthalmologists and technicians with standardized medical equipment. Automatic diabetic retinopathy classification using artificial intelligence and computer vision techniques mostly overcome the above-mentioned limitations despite it still being a challenging task. In this study, pretrained CNN models are applied via transfer learning for diabetic retinopathy classification using retinal fundus images. The MESSIDOR, MESSIDOR-2 and DDR datasets are used in this study. VGG16, InceptionV3 and MobileNet architectures are used for binary classification via transfer learning using the MESSIDOR and MESSIDOR-2, and the highest 84% accuracy is achieved by InceptionV3. The performance of two lightweight CNN models: MobileNet and MobileNetV2 are evaluated for binary and multiclass classification using the DDR dataset. The MobileNet model performed well than Mobilenetv2 in binary and multiclass classification. MobileNet achieved 80% and 71% accuracy whereas mobileNetV2 shows 79% and 69% in binary and multiclass classification, respectively.