Automatic Detection Of Neovascularization In Retinal Images

Diabetes Mellitus is a disease worldwide leading to many complications. Diabetic Retinopathy (DR) is a serious complication of diabetes caused by the damage of blood vessels due to hyperglycaemia in blood. Diabetic retinopathy can present a series of lesions in the retina, such as micro-aneurysms, hemorrhage and exudates. Diabetic retinopathy is the leading cause of blindness among people of working age in developed countries. Diabetic retinopathy should be diagnosed as early as possible for early treatment, and retinopathy screening is an efficient and cost saving way for early diagnosis.The lesions of diabetic retinopathy can be diagnosed using color retinal images captured by fundus cameras for retinopathy screening. In several types of lesions, neovascularization is a key sign for dividing diabetic retinopathy into Non-Proliferative Diabetic Retinopathy (NPDR) and Proliferative Diabetic Retinopathy (PDR), so neovascularization detection is very important for making diagnosis for diabetic retinopathy.In this thesis, a filter bank for analyzing neovascularization textures in color retinal images is presented. The filter bank includes the RGB Channels, Standard Deviation Filters, Anisotropic Gaussian Filters, Gabor Matched Filters and Differential Invariant Filters. This filter bank is invariant both to translation and rotation, and can work in multi-scale.An automatic framework for neovascularization detection is developed in this thesis. The framework contains preprocessing, feature extraction, classification and feature selection. Preprocessing removes the uneven illumination in retinal images, and then the filter bank is applied to the images to extract the features. Extreme Learning Machine (ELM) is used as the classifier in this works for its high performance. For reducing the calculation complexity, a feature selection algorithm is used to select the most import features.Results show the filter set can extract the textures of neovascularization regions in color fundus images, and the framework can detect and mask the neovascularization regions.