Abstract

We propose a robust and accurate method to extract thecenterlines and scale of tubular structures in 2D images and3D volumes. Existing techniques rely either on filters de-signed to respond to ideal cylindrical structures, which loseaccuracy when the linear structures become very irregular,or on classification, which is inaccurate because locationson centerlines and locations immediately next to them areextremely difficult to distinguish. We solve this problem by reformulating centerline detec-tion in terms of a regression problem. We first train regres-sors to return the distances to the closest centerline in scalespace, and we apply them to the input images or volumes.The centerlines and the corresponding scale then corre-spond to the regressors local maxima, which can be easilyidentified. We show that our method outperforms state-of-the-art techniques for various 2D and 3D datasets.