Algorithm Research On Segmentation And Centerline Extraction Of DSA Cerebrovascular Images

Abstract:Cerebrovascular disease has been one of the main diseases threatening human health and life. For the DSA (Digital Subtraction Angiography) images can present the morphology, structure and pathological changes of cerebrovascular actually, it is regarded as the golden standard in diagnosis of cerebrovascular disease.However, the problems of blur overlapping vessels existing in DSA images may caused misdiagnosis and miss-diagnosis. To solve these issues, apply various post-processing algorithms to DSA images is required so that can get a right description of the vascular structure. The segmentation and centerline extraction of cerebrovascular are a significant part during the process:the results obtained not only can be used for3D reconstruction of the vascular tree, may also help doctors to estimate the severity of vascular disease.In this dissertation, several technology problems of vascular segmentation and centerline extraction are discussed.First, a brief summary of the commonly used vascular segmentation algorithms was made and the Fuzzy C-means(FCM) clustering algorithm was mainly described. In order to solve the defects of more iterative times and noise sensitivity, clustering criteria and spatial information were introduced to improve the objective function of traditional FCM algorithm for DSA cerebrovascular image segmentation. The experimental results show that the improved method reduces the iterative times and is robust to the noise. And the segmentation accuracy of the improved method is higher than that of the traditional FGM algorithm.Then, the commonly used vascular centerline extraction algorithms were summarized and classified into two groups:indirect method based on the result of segmentation and direct method based on the original image. After this, two indirect methods were discussed:1) fast marching method (FMM) and distance field combined automatic extraction algorithm;2) refinement algorithm based on index table. The experimental results show that the former method can extract a smooth centerline automatically while the latter one can get centerlines of the whole vascular tree.Finally, the application of minimum cost path (MCP) algorithm which belongs to the direct method finding the centerline was discussed. Applying the traditional MCP algorithm to find the centerline through bend vessel yields a biased path which closes to the one side of vessel wall. To solve this problem, a centering method based on the Gaussian profile is proposed to correct the path obtained by MCP algorithm. And a smoothed centerline was obtained by using gradient and cubic b-spline fitting to the corrected points. The experimental results show that this improved method is robust to noise and the extracted centerline is closer to the center of vessel in the areas of high curvature.