Title:Early Detection of Diabetic Retinopathy and Severity Scale Measurement: A Progressive Review & Scopes

Abstract:Early detection of diabetic retinopathy prevents visual loss and blindness of a human eye. Based on the types of feature extraction method used, DR detection method can be broadly classified as Deep Convolutional Neural Network (CNN) based and traditional feature extraction (machine learning) based. This paper presents a comprehensive survey of existing feature extraction methods based on Deep CNN and conventional feature extraction for DR detection. In addition to that, this paper focuses on the severity scale measurement of the DR detection and to the best of our knowledge this is the first survey paper which covers severity grading scale. It is also necessary to mention that this is the first study which reviews the proposed Deep CNN based method in the state of the art for DR detection methods. This study discovers that recently proposed deep learning based DR detection methods provides higher accuracy than existing traditional feature extraction methods in the literature and also useful in large scale datasets. However, deep learning based methods require GPU implementation to get the desirable output. The one of the other major finding of this paper is that there are no obvious standard severity scale detection criteria to measure the grading. Some used binary class while many other used multi stage class.