Classification of diabetic retinopathy through texture features analysis

Diabetic retinopathy is one of the complications of diabetes that can cause blindness. Early detection is useful to reduce the risk of blindness. There are two approaches of early detection in diabetic retinopathy i.e. lesion characteristics and texture features. Both approaches have advantages and disadvantages. In this study, we use texture feature because is easier to implement. Texture features used in this study is Local Binary Pattern (LBP) because it has better data representation than other algorithms. However, it still needs to be improved. We proposed modified LBP that change paradigm of center point comparison. k-Nearest Neighbor (k-NN) and Support Vector Machines (SVM) was chosen as classifier. We do two scenarios for classification, that is normal-abnormal classification, and four-phases classification. First scenario classifies images into normal and abnormal, while second scenario classifies the image into normal, mild, medium, and severe in disease. As a result, the proposed methods show better accuracy compared to other method. The accuracy for all scenario tested is about 90%.