Graphs are useful representations of the liver vasculature. They support tree matching algorithms in landmark-based registration algorithms, they are useful to separate connected vessels from two different vessel systems and are the basis of vessel annotation tools. In this paper, we propose a hierarchical decomposition of vessel skeletons into sub-branches. This simplifies the process of creating labeled graphs and extracting features. Furthermore, we propose a measure to classify voxels as branch voxels. We applied our method to several datasets with satisfying results and found that the number of sub-branches is normal distributed under rotation.