Retinal Vessel Segmentation Based On Hessian Matrix And Multiscale Analysis

Many diseases can change the shapes and the structures of the retinal vessels. The im-mediate analysis and diagnosis for retinal images carried out by ophthalmologist can preventthe disease from occurring. The early detection and diagnosis of retinal pathological struc-tures is extremely important, while, manual detection for retinal images is a time-consumingand trivial task. So it is popular to make quick segmentation for retinal images with CAD tohelp ophthalmologists to realize the detection.Though many algorithms about the automatic segmentation of retinal blood vesselshave been published, for the reasons of poor contrast, various widths of retinal vessels andthe disturbance of the pathological changes, there are still many challenges in the domainof retinal vessel segmentation. For resolving the problems of consuming time and poor seg-mentation performance on small vessels and broken vessels, we propose an unsupervisedmethod of retinal vessel segmentation. Based on the analysis of the vessel shape, we useHessian Matrix to measure the linear property of retinal vessels.The traditional usage ofthe Hessian Matrix was to detect the vessel centerlines,however,not by using the past usagemode the method presented here gets the eigenvalue map through the Hessian Matrix. Inthis method, we first obtain the coarse retinal vessel segmentation results using the multi-scale eigenvalue map.Then the seed points and undetermined blood points of retinal vesselsare extracted from the previous coarse results, which is followed by tracking these undeter-mined blood points from seed points and analyzing them according to the information ofthe eigenvectors and magnitudes. At last, we make use of the morphological operations forfurther processing to obtain the final vessel segmentation.Our algorithm is experimented on two publicly available databases and the perfor-mance is compared with other methods. Our method adopts the unsupervised methodavoiding the heavy work on labeling the image which appears in the supervised methods.Additionally, our method is simple and effective, takeing the advantage of low computa-tional cost, and what’s more important is that it works well on detecting small vessels.