Comparative performance of deep learning architectures in classification of diabetic retinopathy

This paper analyses the performance of deep learning architecture for classifying the retinal fundus images on diabetic retinopathy (DR) and tracing the severity levels of it. Presently, for classifying these fundus images, many deep learning models are employed with the help of several classifiers. The drawback of several deep learning systems is less efficient output, even in some cases, the wrong classification can be encountered. Since this is medical image classification, utmost care and response have to be given to ensure the proper and exact classification without much complexity. This paper aims to analyse the classification performance of different deep learning architectures with respect to the classification of DR severity levels. From these studies, the concept of CNN algorithms, other transfer learning approaches, and CNN-based models with their dedicated usage for image classification applications especially in retinal fundus image classification was analysed. We have utilised the IDRiD challenge dataset and a custom dataset from a leading hospital to demonstrate image classification using different deep-learning architectures.