Research On Detection Method Of Retinal Features

The structure and the lesion area in the retina contains abundant information that is closely related to human health. Microaneurysm(MA) is a main disease in early diabetic retinopathy, microaneurysms detection is beneficial to the diagnosis and treatment of this disease. The optic disc(OD) is considered as the most obvious structure and starting site of blood and lines of a retinal image. Reliable MAs detection and OD detection is very important for analysis of retinopathy and grasp of fundus.In this paper, we focuses on the contrast enhancement, candidate point extraction and classification for the MAs detection. In the process of contrast enhancement, the factor is determined by local standard deviation and global standard deviation, It can effectively enhance the region with weak contrast and control the region with strong contrast. We proposed a method for candidate point extraction using three layer filter with two value threshold and a direction parameter. The result of candidate point extraction is wonderful with high sensitivity. We use a method for classification based on KNN algorithm, but the number of MAs in one's neighbors decides whether it is MA or not. The method was evaluated in ROC datasets, and the result were satisfactory.In this paper, we proposed a method for the OD detection in retinal images using high-level feature based on deep learning. A feature combining deep feature and traditional handcrafted feature is designed to establish a double layer detection model. The model is composed of two SVM classifier. The first one is to find the candidate region of OD; the second one is to precisely locate the center of OD. The proposed method was evaluated in four publicly available datasets, including the DRIVE, DIARETDB0, DIARETDB1 and ROC datasets. The OD localization accuracy were 100%, 98.46%, 100%, 100% respectively.