Red Profile Moments for Hemorrhage Classification in Diabetic Retinal Fundus Images

Abstract

Reliable and accurate detection of hemorrhages is essential in screening of diabetic fundus retinal images. The hemorrhages are often irregular in shape. They may occur without any other signs of Diabetic Retinopathy. In this work, we captured the irregularity itself as a feature of hemorrhages. We define profile of segmented region and used its centralized moments as novel feature. The Support Vector Machine and its variants are explored for classification. The experiments show promising results with moments of red channel profile. The accuracy of 98.09% is achieved at lesion level and 100% at image level with RBF SVM. The approach with profile moments possibly can be further experimented further for other irregular object detection.