Detection of Hard Exudates Using Evolutionary Feature Selection in Retinal Fundus Images

It is one of the most vital symptoms of DR (diabetic retinopathy) called hard exudates (HE), which are the leakage of cellular debris and lipoprotein from damaged blood vessels of retina. The vision loss is avoided if the detection of HE in the beginning times. Therefore, a novel method is proposed to detect hard exudates automatically. Previously, for exudate prediction supervised and unsupervised methods have been used. Fault detection of hard exudates, miss classification rate will affect these models because of the characteristics like, similarities with other components in the retinal image and intra variations. For that, the retinal fundus images has been used as input. Then these images are pre-processed with some pre-processing algorithms like image enhancement, equalization of histogram to improve the proposed system performance. Total image data files are divided to training and testing datasets. Features are extracted for training and testing using feature extraction algorithm individually. Then classifier algorithm predicts whether the hard exudate is proliferative or non-proliferative. We obtained accuracy of 99.34% using our proposed methods on public datasets like DIARETDB1 and DRIVE.