Retinal Image Vessel Segmentation Based On Clustering Algorithm

Retinal vascular is the only non-invasive direct observation deeper vessel. It can directly reflect the effect of impaction on the cardiovascular disease, and indicates the severity of the diabetes. The detection and extraction of blood vessel from retinal image is with great practical significance for clinical diagnosis and evaluation criteria. In addition, as the body biometrics, the retinal blood vessel which structure is stable and concealed has a wide application in the security protection and body identification.Based on the investigation of former researchers, the method for retinal image vessel segmentation based on clustering algorithms is studied. The main work in this paper is as follows:(1) The preprocessing of retinal vessel images is studied. Through analysis of red, green and blue channel images from retinal color image, it is found that the image contrast of green channel is the highest, so a method to de-noise and adjust the brightness on the green channel image is presented, then the two-dimensional Gaussian matched filter is used to enhance the retinal image. The simulation results show that the performance of this method is desired.(2) The influence of original clustering center, kernel function, clustering number and weighted index on segmentation result is studied. The parameters of clustering algorithms are selected according to the simulation and evaluation criteria of clustering algorithm. The simulation results show that when fuzzy clustering number is 5, kernel function is Gauss kernel, the weighted index is 2 and using K-means clustering center as original clustering center, this method can get good quality of segmentation result.(3) A novel scheme to extract the retinal vessel area automatically is proposed based on the relation between the segmentation image and corresponding membership degree. Analysis results show that a satisfactory segmentation effect can be obtained while three cluster segmentation images with the smaller membership degree are merged as the vessel area.(4) In order to prevent the fuzzy clustering algorithm based on objective function from falling into local optimization, genetic algorithm is introduced to optimize the objective function value.(5) For the weakness of poor contrast and the background area variance of retinal image, especial the retinal image is interfered by noise seriously, it is difficult to extract the retinal vessel using the single method. A method based on kernel fuzzy c-means clustering and three-dimension OTSU is presented for this situation. Firstly, the main retinal vessel is extracted using the kernel fuzzy c-means method, then the small retinal vessel is extracted using the three-dimension OTSU method coupled with GA. The simulation results show that its performance is satisfied.