定向区域生长算法及其在血管分割中的应用

Directional region growing algorithm and its applications in vessel segmentation

针对医学图像中微细管道结构灰度连续性差,采用常规区域生长法进行分割容易丢失末梢的问题,提出一种定向区域生长算法,可以在生长过程中跨越管道结构中的低灰度区域。算法向图像中已生长区域外灰度最高的方向进行生长,每次将一个体素加入已生长区域,将图像转变为一颗以种子点为根结点的树,再从叶子结点进行回溯以确定感兴趣区域。对实现算法的数据结构进行了讨论。算法可以应用于任意维的图像。对2维和3维图像的测试结果表明,相对于常规的区域生长法,算法可以分割出更多的血管分支。算法对3维图像的运行时间为秒钟量级,可以满足临床应用的要求。

Accurate extraction of the vasculature in medical images is prerequisite to structural analysis and further applications such as surgical planning. Region growing algorithm is a simple and effective method to extract thick blood vessels which makes use of the spatial continuity of the vascular tree, while the extraction result of small vessels like hepatic artery is unacceptable. In order to solve the problem that the continuity of tenuous vasculature is poor in medical images and vessel segmentation based on traditional region growing may lose distal branches, a directional region growing （DRG） algorithm is proposed which can skip the low gray area in the vaseulaturc during the growing process. The algorithm grows towards the direction of the maximum gray around the grown region, and adds one voxel to the grown region in each iteration. The image is transformed into a tree after the growing process in which the seed point is the root. A trace back procedure beginning from the leaf nodes of the tree can finally determine the region of interest （ROI）. The algorithm relaxes the conditions to determine ROI, and small area with low gray in the ROI is permitted. There are two timeconsuming steps in the algorithm due to the enormous amount of data in 3D medical images, one is to determine the growing direction in each iteration, the other is to construct the paths from the seed point to leaf nodes during the trace back procedure. Data structure to improve the speed of the algorithm is discussed. The algorithm can be applied to images with any dimension. The algorithm is tested with 2D and 3D images. In both conditions, the segmentation results obtained by DRG contain more distal branches in comparison with tractitional region growing algorithm. To some vein phase CT images with poor quality, the proposed algorithm can also generate better results. Four parameters should be appointed in the algorithm and the empirical values are given. The computational time of the algorithm on 3D images is several seconds, which is acceptable in clinical applications. The surface of the extracted vasculature is rough due to the discrete nature of digital images, and further study is needed to smooth the surface before visualization.