Automated method for hierarchal detection and grading of diabetic retinopathy

This paper proposes an automated method for comprehensive detection and grading of non-proliferative diabetic retinopathy (NPDR) taking all DR abnormalities into consideration. Abnormalities of NPDR like red lesions and bright lesions are detected in chronological order according to medical information. This strategic hierarchal sequence of abnormality identification of these pathologies provides an opportunity to have an efficient and computationally optimised method for diagnosis of DR. Unlike previously reported methods, this proposed method includes both bright and red lesions for detection and grading of NPDR. Red lesions and bright lesions are detected using adaptive threshold that uses local statistical features from the fundus image for segmentation of DR abnormalities which is invariant to the quality of the image and noise content. Experimental results show that the proposed system achieves classification sensitivity/specificity for bright and red lesions as 97/89% and 94.2/84.5%, respectively. Classification of normal images from DR images achieves an average sensitivity/specificity of 93.90/76.49%.