Hierarchical detection of hard exudates in color retinal images

Diabetic retinopathy (DR) is one of the common causes of blindness, and hard exudates (HEs) are the primary and early clinical signs of DR. Thus, a reliable detection of HEs is significant for clinical diagnosis and preventing vision loss of patients. In this paper, a novel method is presented to detect HEs automatically in color retinal images. The method consists of two stages: coarse level and fine level. In coarse level, we extract HEs candidate regions by combining histogram segmentation with morphological reconstruction. While in fine level, we define 44 representative features for each candidate region, and train a support vector machine (SVM) model to classify HEs and non-HEs. We evaluate the proposed method on the public DIARETDB1 database and yield a sensitivity of 94.7% and a positive predictive value of 90.0%. Experiment results show that our method can detect HEs efficiently.