BPSO-Based Feature Selection for Precise Class Labeling of Diabetic Retinopathy Images

Abstract

Diabetic retinopathy (DR) is an eye disease caused by high levels of blood sugar in diabetic patients. A timely detection and diagnosis of DR with the aid of powerful algorithms applied to the eye fundus images may minimize the risk of complete vision loss. Several previously employed techniques concentrate on the extraction of the relevant retinal components characteristic to DR using texture and morphological features. In this study, we propose to estimate the severity of DR by classification of eye fundus images based on texture features into two classes using a support vector machine (SVM) classifier. We also introduce a reliable technique for optimal feature selection to improve the classification accuracy offered by an SVM classifier. We have applied the Wilcoxon signed rank test and observed the p-values to be 3.7380 × 10−5 for Set 1 and 7.7442 × 10−6 for Set 2. These values successfully reject the null-hypothesis against the p-value benchmark of 5%, indicating that the performance of SVM has significantly improved when used in combination with an optimization algorithm. Overall results suggest a reliable and accurate classification of diabetic retinopathy images that could be helpful in the quick and reliable diagnosis of DR.