Automatic detection of diabetic retinopathy from fundus images using image processing and artificial neural network

Abstract

Diabetic Retinopathy (DR), a disease that affects the retina, is a complication of diabetes and the fifth most common reason causing blindness and loss of vision over the world. The risk of severe damage to vision can be reduced by early screening and treatment. The absence of any visual symptom, lack of an adequate number of ophthalmologists, difficulties in the manual detection process, etc. are the key challenges for early detection of DR. An automated detection method would address these issues and facilitate the detection process. DR can be diagnosed by analyzing pictures of retina (known as fundus images) captured by special cameras. The severity of DR depends on the properties of some anatomical features present in a fundus image namely - optic disc, blood vessel, exudate, microaneurysm, and hemorrhage. This work has first developed techniques to identify each of the components separately from a fundus image. In the course of that, it has been able to overcome many limitations prevailing in state-of-the-art techniques. After that, a technique for detecting the severity of DR has been developed based on the identified components.