Fully Automatic Hepatic Portal Vein Segmentation

Hepatic vasculature analysis and visualization are very essential in preoperative surgery planning. Surgeons must have a deep understanding of intrahepatic vessel systems in order to determine a proper hepatic resection approach and avoid unnecessary massive hemorrhage caused by main vessel rupture in surgeries.Traditional hepatic vessel segmentation methods were mainly based on region growing method and thresholding method. However, due to strong image noises in CT images, their segmentation results were unpromising. Besides, they also suffered the other two problems:lack of robust methods to automatically determine the initial parameters and to separate falsely connected portal system and hepatic venous system.The aim of this paper is to provide a robust and fully automatic vessel segmentation method to extract the hepatic portal system and separate it from the hepatic venous system. Firstly, our method utilizes anatomical priors to roughly locate the portal vein trunk (PVT), and adopts histogram analysis to precisely extract it as the initial area, thereby automating the initialization of portal vein segmentation. Secondly, a novel vessel segmentation framework --- Propagations of Vessel Region on Slice (PVRS) is proposed and used with a vesselness measure derived from multiscale analysis of tubular structure to segment hepatic vasculature. This framework is anisotropic, which divides 3D propagation into inter-slice and intra-slice one and thus quite suitable to deal with anisotropic volume data such as CT data. Finally, graph analysis on the vascular skeleton is used to remove noisy end branches, loops and falsely connected hepatic veins from the portal system based on our pruning rules.The computation of vesselness is quite time consuming. To accelerate its computation, we exploit the spatial locality of the framework and propose a way to adaptively determine the scale range of each voxel. The experimental results show that the optimized computation speed is roughly 18 times faster than the original one. In addition, we also compare our segmentation results with that of intensity-based region growing method and find our method achieves obviously better results.