Blood Vessel Segmentation and Classification of Diabetic Retinopathy Images using Gradient Operator and Statistical Analysis

Retinal blood vessel detection in fundus images is a challenging task. Widely spread blood vessels in diabetic retinopathy effected fundus images can not be accurately segmented using conventional gradient-based edge detection techniques. Accurate detection of blood vessels can give better classification for decisive diagnosis at different stages of diabetic retinopathy. The proposed blood vessel segmentation technique is a combination of gradient and morphological operators. The fundus images are preprocessed using a local phase-based enhancement technique to highlight the blood vessels from the background image. Further, detected edges are made connected using an averaging filter. Area occupied by the blood vessels provides a biomarker to classify fundus images into normal, prolific diabetic retinopathy (PDR), and non-prolific diabetic retinopathy (NPDR). The classification is achieved using ANOVA test. The ANOVA test classified with the 91% accuracy for normal, 92.7% for PDR, and 87.8% for NPDR images, respectively. These results are compared with conventional Canny and Prewitt edge detection techniques and the proposed method outperforms both.