Detection and identification of hemorrhages in fundus images of diabetic retinopathy

Diabetic retinopathy(DR)is a progressive eye complication of diabetes which may cause sight loss and blindness if not detected and treated in time. Hemorrhage is one of the early signs of diabetic retinopathy. The accurate detection of hemorrhages is critical for early screening and diagnosis of DR and protecting the vision of diabetic patients. In this paper, we present a novel method that detects and identifies hemorrhages of diabetic retinopathy in fundus color images. Firstly, the fusion algorithm of morphological top-hat transform and maximum entropy thresholding are applied for enhancement and segmentation of blood vessels. This, in turn, captures thin vessels and produces accurate vesselness measures in low-contrast and local intensity of DR images. A feature vector is formed by extracting the size, shape, intensity and statistics of the hemorrhage candidate regions. The extracted feature vector is then used in a Fuzzy C-means clustering classifier to exclude spurious hemorrhages. Finally, the proposed method is evaluated and tested on the two standard diabetic retinopathy datasets DIARETDB0 and DIARETDB1. Experimental results show that a mean sensitivity of 92.87% and a mean positive predictive value of 88.62% are achieved by the proposed algorithm. Further, the experimental results also demonstrate that the method described herein has higher sensitivity and positive predictive value than recently published methods, which make the proposed scheme better suited for early diagnosis and treatment of DR hemorrhages.