Automatic Classification of Diabetic Retinopathy Based on Deep Learning - A Review

Abstract

India is home to approximately 70 million people with diabetes, and this epidemic is estimated to increase to 130 million by 2045. Diabetic retinopathy (DR) is a dangerous eye condition that affects diabetic persons. DR remains asymptomatic until vision is affected. Treatment is most likely to be effective when performed before progression to advanced disease. Therefore, early diagnosis of DR is crucial for its treatment as it can eventually cause permanent blindness. DR is the most common complication of diabetes and about 3 to 4.5 million people in India suffer from vision threatening DR. The treatment requires costly devices and medications and the disease requires regular follow-up from diagnosis to the end-of-life. Also, manual inspection of fundus images by experienced ophthalmologists to check morphological changes in microaneurysms, exudates, blood vessels, hemorrhages, and macula is a very time-consuming work. It is also subject to substantial inter-observer and intra-observer variability. Many deep learning algorithms are being used currently which can perform automatic classification of images that are input to the system. This paper is a review of deep learning techniques applied for the detection and classification of DR using retinal images. The factors that may influence the performance of a deep learning algorithm in detecting DR are also considered.