A Comparative Study on the Various Neural Network Approaches to Classify Diabetic Retinopathy

Abstract

Diabetic Retinopathy affects more than 85% of patients with longstanding diabetes and is it is one of  the  primary causes of blindness for the age group 20-64. DR can     be divided into two types: non-proliferative diabetic retinopathy (NPDR) and proliferative diabetic retinopathy (PDR). In recent years, more than 85% of people suffer from diabetic retinopathy due to lack of proper diagnosis or early prediction. Several methodologies have been proposed involving various core con- cepts to diagnose this issue. The present model  classifies  DR  into five categories with integer values from zero to four. Deep CNN approach proves to be sufficiently efficient but defining   the problem uniquely in  order  to  ensure  the  occurrence  of  the disease still remains a problem. Even though CNNs are popular for their generally high accuracy rate, there are obstacles like computational complexities and processing time. Region Proposals are a way to solve this problem. Using Region based CNN approaches, the region of interest for the purpose can be detected. Another challenge is the unavailability of a universally or majorly accepted  database  of  fundus  images  which  makes it harder to get accurate results for the algorithms. One such database, the DRiDB proposes to overcome this obstacle. In this paper we try to present a comparative study of the various region based proposals and object detection methods that can be used   to predict diabetic retinopathy and make it more accurate and efficient.