Retinal Neovascularization (NV) is one of the most important causes of diabetic blindness which emphasizes the serious need for automatic screening tools to diagnose diabetes and treat it as early as possible. In this work, the problem of NV detection in the color images of retina is considered. Previous researches in the analysis of retinal images either ignore such complex lesion or in few cases approach it superficially and pose many limitations. Obtaining a complete vessel map and then trying to identify NV is a recent methodology. Unfortunately, NV mainly affects small and hard to detect vessels, hence, we believe such an approach is inherently limited. Therefore, in this paper six different texture descriptions are proposed as features in the classification of retinal images into NV and normal group. We study application of the proposed features using a support vector machine (SVM) as a simple classifier to emphasize the importance of features against the complexity of the classifier. Feature extraction is performed on local regions of the images (128×128) to localize the lesion. A dataset of 4541 regions is considered that contains 1319 NV regions which are manually marked by an expert. The results show the proposed texture descriptors are able to reveal the NV lesion with an acceptable accuracy of around 90%.

Keywords