Improved Sparse Coded Features for Automatic Identification and Discrimination of Exudates and Drusen in Retinal Fundus Images

Diabetic Retinopathy is a medical condition in people having diabetes in which damage occurs to the retina and can cause partial or complete vision loss. It is the most common reason for vision loss in the world. Diabetic Retinopathy is characterized by the appearance of retinal lesions such as microaneurysms, exudates, cotton wool spots, and haemorrhages. Age-Related Macular Degeneration (ARMD) is another eye disease which causes gradual loss of vision with age and is characterized by the appearance of drusen. Exudates and Drusen are known as bright lesions as they appear as yellow deposits in the retina. In this paper, we propose an automated method for the classification of bright lesions in retinal fundus images based on improved sparse coded features and Support Vector Machine (SVM). We employ Scale-Invariant Feature Transform (SIFT) to extract interesting regions in the image in the form of key points and extract local low-level features in the neighbourhoods of key-points. After that, we use the sparse coding technique to get high-level sparse coded features that are used for classification using SVM. Experimental results reveal that the proposed approach gives an accuracy, sensitivity, a specificity of \(92.90\%\) $92.90\%$ , \(89.60\%\) $89.60\%$ , \(95.40\%\) $95.40\%$ for exudate classification and \(93.80\%\) $93.80\%$ , \(89.60\%\) $89.60\%$ , \(95.80\%\) $95.80\%$ for drusen classification, with a dictionary of 300 atoms on a dataset of 1158 images.