Extraction of Microaneurysms and Hemorrhages from Digital Retinal Images

Abstract

Detection of red lesions from color fundus images is crucial in the early detection of diabetic retinopathy. Automatic red lesion detection is a challenging task as they have low contrast, irregular shapes, variable sizes and resemblance of their intensities with blood vessels. This paper presents a novel hybrid red lesion detection system that combines phase congruency based and mathematical morphology based methods to detect candidate red lesions. The significant contribution of this paper is the computation of phase congruency using extended 2D log gabor filter. The proposed red lesion detection system is a three stage system which combines polynomial contrast enhancement for preprocessing, hybrid detection for coarse candidate red lesion extraction, and support vector machine classifier for fine segmentation of red lesions. Experimental evaluations of the proposed system using publicly available fundus image databases demonstrates superior performance over other red lesion detection algorithms recently reported in the literature.