Research On Retinal Blood Vessels Segmentation

The retinal vasculature is the only part of the body’s circulatory system that can be observed non-invasively by optical means. A large variety of diseases affect the vasculature in a way that may cause geometrical and functional changes. Retinal images are therefore not only suitable for investigation of ocular diseases such as glaucoma and cataract, but also for systemic diseases such as diabetes, hyper-tension and arteriosclerosis. Moreover, the retinal vasculature can be used to personal identification as it is unique, stable, occult and mineable. This makes the retinal vasculature a rewarding and well researched subject on clinical medicine and personal identification, security and secrecy.Nowadays, researches at home and abroad make a lot of studies for segmentation of retinal vasculature. This thesis refers for a large number of literatures about retinal blood vessel segmentation and digital image processing, makes a summary to the present achievement, and find that it is unsatisfactory to segment the small blood vessel, especially when the gray contrast is low. To solve this problem, this thesis proposes two retinal blood vessel segmentation methods based on tracking and window, makes a lot of simulation experiments based on the REVIEW database and DRIVE database, calculates the accuracy, sensitivity, specificity and compares with the results of manual and existed methods.(1) The method based on tracking:this thesis proposes a retinal blood vessel segmentation method based on Hessian matrix and local tracking. Firstly, enhance the image use Hessian matrix, then tracking the blood vessel with gradient vectors field of enhanced image under multi-direction. Finally, get the final result of retinal blood vessel by fusing the results of all orientations. Experiment results show that, this method can segment more small vessels, and its accuracy is 0.93871 and 0.93568 in REVIEW database and DRIVE database, better than the classical methods. At the same time, this method improves the sensitivity and specificity, plays a better performance on the segmentation results.(2) The method based on window:this thesis proposes a retinal blood vessel segmentation method based on morphology and Cake filters. Firstly, preprocess the image by using the top-hat translation, then filter the image by Cake filters under multi-direction, and fuse the real part of orientation scores of all orientations. Finally, binary the fusion image adaptably to get the segmentation result. Experiment results show that, this method’s accuracy is 0.95905 and 0.953785 in REVIEW database and DRIVE database, plays a better performance on the accuracy, sensitivity and specificity than existed methods. Moreover, this method can segment more small vessels, especially for the small vessels where the gray contrast is low.