Computer-Aided diagnosis systems for Diabetic Retinopathy: A comprehensive review

We present a systematic review of prominent methods for computer-aided detection of Diabetic Retinopathy (DR). We evaluated the recent methods for DR lesion detection with study of validation dataset types, number of training samples in dataset, sample resolution, performance in terms of accuracy and techniques for classification. Our review reflects that most of the current automated DR detection systems have been validated on datasets with very small number of retinal images. Moreover, it is found that the machine learning (ML) / deep learning (DL) DR detection methods have achieved the classification accuracy up to 100% therefore demonstration of superior performance in terms of accuracy improvement has limited scope. Some DR grading methods have reported accuracy up to 100% however they have limitation in terms of generalization ability as they have been validated on single dataset.