Improving curve skeletons of tubular volumes

Curve-skeletons encode both geometrical and topological information of 3D volumes (sets of voxels), and are key to many applications. However, due to the complexity and variability of the shapes, there is a variety of algorithms yielding skeletons suitable for certain objects, but inappropriate for others. In this article, we are interested in filtering skeletons of digital tubular objects with varying-diameter and junctions, resulting from the segmentation of organs (airway-trees, vessels). Existing skeletons might not be centered or contain faulty branches. For medical applications such as virtual endoscopy or cross-section estimation, post-processing steps are applied in order to prune the spurious branches and smooth the skeleton. State-of-the-art methods for pruning are insufficient with respect to varying-diameter tubes. We propose a new approach to prune irrelevant skeleton branches, and to recenter the skeleton inside the shape.