Curve Evolution Models And Application To Retina Image Segmentation

Vascular disease such as hypertension and heart disease are common disease that damage human’s health, and vascular imaging plays a significant role in diagnosis and treatment of vascular disease. With the development of imaging technology and improvement of imaging resolution, the modern imaging equipment generates so huge amount of data, that makes manual interpretation unrealistic. Computer-aided diagnostic techniques based on image processing are an effective way to solve this problem. In this paper, we take advantage of the curve evolution model which combines the upper prior information and low level image information to study segmentation of retinal blood vessels, which has the characteristics of small diameter, complex topology and inhomogenity. The study could meet the needs for object targeting and quantitative evaluation. The main contents are as follows:Because the traditional parametric curve evolution model demands strict initialization, and the balloon force model can not change the evolving direction, this paper presents a new external model based on regional density probability distribution function. By exploiting the density distribution information which is ignored in traditional Snake model and balloon Snake model, we establish the parametric probability function inside and outside. The curve evolves in certain direction with certain velocity by checking the probability that it belongs to different regions simultaneously, which could increase the capture range of external force and avoid ’leakage’Owing to the advantage that the geometric curve evolution could change topology, we study the region based curve evolution model. In order to balance the contradiction of scale selection and accuracy of segmentation while segmenting the long and thin shaped blood vessel whith improved LBF curve evolution model, we utilize the local intensity information and extend the dimension of the distance function to alleviate the above problem.To meet the requirements for segmentation of retinal blood vessel and optic nerve head, we utilize the curve evolution model discussed before to segment the two tissues:1) when segmenting retinal blood vessel in retinal images with complex and inhomogeneous background, we use Gabor transform to enhance blood vessels to facilitate the subsequent processing. As the magnitude of gradient varies in different boundary region, we present a more regurarized term based on vector flux flow and then the improved LBF model combined with the proposed item is used to segment blood vessel.2) considering the difference of the shape characteristic between the optic nerve head and blood vessel, we present a round shape constrained curve evolution model to segment optic nerve head even when it lacks some information in images.Experiments show that the proposed method could segment blood vessels and optic nerve head in retinal images with high accuracy even when vascular diameter varies, the shape is complex, the background is inhomogeneous.