Classification of Colored Retinal Images for Diabetic Retinopathy Based On HOG Feature Selection.

Abstract

Diabetic Retinopathy is an eye disease that affects the people suffering from diabetes. The high sugar levels in blood leads to damaged blood vessels in eyes. Diabetic retinopathy is one of the major causes of blindness in present day. Diabetic retinopathy is mainly identified by red spots known as micro-anuerysms and bright lesions called exudates. It has been seen that the early detection of diabetic retinopathy is done mainly by identifying these exudates. Therefore an automated early detection of diabetic retinopathy is need of the moment. Inspired by this manifestation of the exudates in the eyes of person suffering from diabetes, a framework is proposed for the early detection of diabetic retinopathy. In this research we consider the colored retinal images of the eyes taken by fundus camera. The proposed system performs the feature extraction on the retinal images after the preprocessing stage. The features are extracted by HOG technique. Finally, the classification for normal and abnormal retinal images is done by using well trained KNN, SVM and Random Forest. Ensemble model of the three classifiers namely, KNN, SVM and MLP is also used for achieving a better accuracy. This approach is evaluated on 400 colored fundus images, including two publically available dataset. The accuracies of 89%, 87%, 95% and 93% are obtained for the three individual classifiers and the ensemble model respectively.