No abstract

Diabetic retinopathy (DR) is the most common diabetic eye disease and a leading cause of blindness. It is an irreversible loss that needs to be detected early to save the patient's vision. For early detection, premature symptoms of DR can be noticed in exudates, blood vessels, micro-aneurysms, and hemorrhages of the retinal eye surface. Thus, detecting grades of DR from individual features is an active area of research. This work aims to extract features from retinal images and classify DR to binary and multiclass classification. To detect different grades of DR, various machine learning classifiers are applied on individual and combined features. Deep learning models, i.e., convolution neural networks (CNNs), and its variants are also used to diagnose and classify the severity of DR accurately in multiple classes. A fine-tuned variant of VGG-19 model is introduced which uses the transfer learning approach and is evaluated on digital fundus retinal images for both binary and multiclass classification. Experimental results revealed that the fine-tuned model of VGG-19 performs better than contemporary CNN models and machine learning algorithms. On average, it improves by 37% from existing CNN models on multiclass classification. The work presented here provides an ophthalmologist the comprehensive way to detect the grades of diabetic retinopathy even from single feature only.