The Fusion of Features for Detection of Clinical Symptoms of Diabetic Retinopathy and its Grading from Digital Fundus Images

Abstract

Diabetic Retinopathy is a recurrent retinal disorder that can affect patients with both type 1 and type 2 diabetes. From a large amount of data available, it is apparent that Diabetic Retinopathy is a progressive disease that can terminate in severe vision problems or complete vision loss. The clinical symptoms of Diabetic Retinopathy include microaneurysms, haemorrhages, hard and soft exudates. They are lesions seen on the surface of the retina among diabetic patients. They indicate a pre-proliferative Diabetic Retinopathy state and needs to be treated. This paper describes a method based on morphological operations and thresholding for detection of clinical symptoms of Diabetic Retinopathy and grading them. As a part of the feature extraction stage, feature level fusion comprising LBP and HOG features is explored as it yields better accuracy in detection of affected areas in fundus images. Various classifiers namely; SVM, k-NN, Decision tree and ECOC classifiers have been tested. The average accuracy values for detection of microaneurysms and hemorrhages using fusion features yielded 95% for SVM, 97% for k-NN and 96% for Decision tree classifier respectively and the average accuracy values for detection of hard and soft exudates using fusion features yielded is 98% for SVM, 91% for k-NN and 98% for Decision tree classifier respectively. The proposed method using fusion features for grading of Diabetic Retinopathy yielded an accuracy of 32% for k-NN, 91% for Decision tree and 98% for ECOC classifier respectively. The methods and algorithms developed as a part of this research work will aid the ophthalmologists and clinicians in early detection of retinal disorders. It will also be useful for mass screening programs particularly in rural areas and developing nations.