Deep Learning Techniques for Early Detection of Diabetic Retinopathy: Recent Developments and Techniques

Abstract

Diabetic Retinopathy (DR) is a health disorder in human retina, caused as a result of Diabetes Mellitus (DM). It leads to loss of vision and in severe cases it results in blindness, as a result of mutilation of the retina. Statistical data estimates that 80% of diabetic patients, suffering from prolonged diabetes, also suffer from DR. Hence, in the present time DR has become an imperative matter and requires primary stage evaluation and assessment such that loss of vision and blindness can be averted. However, the physical diagnosis of the disease is laborious and susceptible to error. Besides, the convenience of availing Ophthalmologist irrespective of place and time, is not possible. Thus, the necessity of an exceedingly enhanced and computerized intelligent system arises, that can be engaged for the initial stage detection of DR. A number of Machine Learning models are proposed by researchers since decades, for the diagnosis of DR. Various feature extraction techniques are also proposed for deriving prominent retinal lesions, for initial stage diagnosis of DR. However, traditional Machine Learning models showcase poor generalization during feature extraction because of smaller datasets. This can be overcome through use of Deep Learning models, larger dataset and high computing processing units for generalization. This paper aims to give an overview about DR, a brief description of the earlier works, and the current automated systems and advancements, for the purpose of early detection of DR. This paper also focuses on the state-of-the-art DR lesions, origin and signs of DR, categories of DR and state-of-the-art Deep Learning models, that are proposed and applied for DR detection, at the preliminary stage.