Segmentation of 3D Tubular Tree Structures in Medical Images (print-version)

The segmentation of tubular tree structures like vessel systems in volumetric medical images is of vital interest for many medical applications. However, a diverse set of challenging objectives and problems is related to this task in different application domains. In this work, we develop and evaluate methods to address these issues. To accomplish the segmentation of heavily branched structures in a robust manner, we propose a generally applicable three-step approach consisting of: (i) a bottom-up identification of tubular structures followed by (ii) a grouping and linkage of these tubular structures into tree structures that are (iii) used as a prior for the actual segmentation. This approach incorporates additional prior knowledge compared to conventional approaches: the individual tubular structures have to be connected with each other and – from a biological perspective – to be supplied. In this way, we achieve a high robustness regarding the structural correctness of the segmentation results. We develop and investigate novel methods for each of these processing steps addressing the needs of different applications. In particular, we present a novel approach for detection of tubular objects using the Gradient Vector Flow to address limitations of the typically used Gaussian scale space. We propose two methods for grouping and linkage of sets of unconnected tubular structures into tubular tree structures. One enables an extraction of high quality centerlines in regions that deviate significantly from a typical tubular shape, while the other one allows for a separation of interwoven tubular tree structures as well as handling of various kinds of disturbances. To accurately segment the identified tubular structures two methods are developed. One solves the segmentation task in a globally optimal way using graph cuts, while the other one segments according to the edge closest to the centerline. Based on these methods, different applications for segmentation of blood vessel trees (liver vasculature and coronary arteries) and airway trees in CT datasets are developed. The methods are evaluated on clinical datasets and compared to results achieved with other state-of-the-art methods developed for the same task. The results successfully demonstrate