Retina Vessel Segmentation And Artery-vein Vessels Classification

As the only microvasculature that can be directly watched without invasive means in the human body, retinal vessel has been paid a large number of attention by the researchers. Physical change of the vessels can be the prediction and representation of many human diseases. A mass of clinical trials indicate that such as diabetic retinopathy have a tight relation with the abnormal of Arterio-Venous ratio, by early obtaining these anomalies will be helpful to slow down the development of the deterioration and even put an end of the happening of the disease. In addition, like high blood pressure, and some of pancreatic disease can cause retinal vascular network anomalies. But due to the complexity of retinal blood vessel structure, relying on the manual analysis of the medical staff can be time consuming and much expensive to the patients, and it is difficult to obtain accurate results. Therefore, adopting the method of image processing to automatic analysis the state of illness has broad prospects.Our major work consists of three parts: retinal images preprocessing, retinal vessel segmentation, and artery-vein classification. We have different strategies to deal with the retinal image to solve different phase images in order to achieve the best result at every stage. We used a modified B-COSFIRE method to segment vessel from the background in fundus images. This method is an unsupervised method and can reduce the dependence on the labeled samples. As for the part of artery-vein classification, we used two kinds of methods. The first method is that we choose various color features in different color channels, and using the SVM to classify the samples. Another method is that applying the deep learning method to the classification, we establish a convolutional neural networks to classify the vessel into arteries and veins. The above two methods need to introduce vessel structure to vote for correcting some wrong labeled samples.By studying the retinal blood vessel arteriovenous characteristics, we achieved an automated system from vessel segmentation to arteriovenous classification. Our method is tested on DRIVE public database and the classification accuracy is 88.7% for pixels and89.07% for vessel lines, respectively, which demonstrate the effectives of our approach.