Research On Vessel Segmentation In Color Fundus Images

Researchers always pay great attention on retinal blood vessels over the years dueto the vessels have an important application value on analysis of retinal vascular lesionsand diagnosing cardiovascular diseases. Under normal circumstances, the structure andmorphology has been in a stable state. However, cardiovascular and cerebrovasculardiseases, such as hypertension, diabetes and coronary atherosclerosis, that seriouslyharm human health, cause changes in the structure of retinal vessel diameter and thedegree of bending. As a result of the retinal blood vessels having complex structure,misdiagnoses often occur when doctors inspect the patients by their naked eye. So thestudy on scientific and effective retinal blood vessels segmentation methods has animportant significance.First of all, we pre-processed the fundus retinal image on the green channel byshading correction and bilateral denoising, and then we extracted the retinal vesselcenterlines with direction-line detection operators in the range of0~180°from12directions. Eventually, we get all the centerlines detected from every direction togetherto form the complete retinal vascular centerlines. Our experiments showed that thecenterlines detected by direction-line detection operators were not only accurate butalso complete.On the basis of the resulting vessel centerlines, the paper used two methods tolocate the boundaries of retinal blood vessels. The first retinal vessels segmentationmethod is based on GrowCut method. We treated the detected centerlines as the seedpixels of blood vessels; the pixels treated by morphology as the seed pixels ofbackground, and then labeled the seeds of vascular and background pixels respectively.Finally, to segment retinal vessels, we assigned the corresponding labels to the residualpixels with cellular automata. This method is fast and can split out most of the bloodvessels, but there are problems that the vessel boundaries are not smooth enough andprecise enough and misclassification phenomenon is obvious in some areas due to thesegmentation process failed to take the structure characteristics of blood vessels intoaccount.In order to use the morphological structure characteristics of blood vesselseffectively, this paper puts forward the second segmentation method: the method based on the improved B-Spline Ribbon Snake to segment blood vessels in the retina image.This method set the initial contour lines which are close to both sides of the vesselcenterlines. We have designed the width energy and region energy in the B-SplineRibbon Snake model, and segmented retinal blood vessels by minimizing the energyfunction. The results of the experiment showed that our method segmented retinalvessel more accurately and smoothly and also segmented low contrast vesselseffectively.