Automated severity scoring for diabetic retinopathy using fundus photography

Abstract

Diabetic Retinopathy (DR) is a micro-vascular complication of diabetes and is one of the major causes for blindness. DR is asymptomatic in early stages and when symptoms surface irreversible damage would have already occurred to the retina hence screening is of at most importance. Here, we propose an automated method for severity scoring of DR which can be utilized for automated filtering and prioritizing patients for ophthalmic review. The proposed method assesses the image quality, detects red lesions and white lesions. The output of the optic disc detection was used to demarcate the retinal image into four regions. Lesion specific features extracted from demarcated regions were used to create a multi-class SVM classifier which in turn was used to arrive at a DR severity score. The proposed method was validated on 829 retinal images and achieved an accuracy of 72.4%. Grouping all the positive grades into single class resulted in a sensitivity of 89.9% and specificity of 77.8%.