| Blood vessels are important features in thoracic CT scans. In this thesis, a general framework for vessel-based analysis of thoracic CT scans is proposed. The framework includes vessel enhancement filters, vessel tree reconstruction, and vessel-based volume registration. In this thesis, two sets of enhancement filters are proposed. One set is based on the correlation matrix of the regularized vectors after splitting. The other set is based on the probabilistic modeling of the distribution of the orientation of the regularized vectors. The proposed filters can distinguish junctions from nodules, including attached nodules, and introduce smaller distortion than existing filters. The utilization of the first-order partial derivatives and the adaptive local windows also makes the proposed filters less sensitive to noise and the local perturbations. The vessel tree reconstruction is based on a novel fuzzy shape representation of vessels and so can cope with structural distortions. The orientations of linear vessel segments are adjusted through homogeneity maximization. Constraints are applied during the merging of vessel segments to improve the robustness of the vessel tree reconstruction. Experiments show that by using the constructed vessel trees, up to 38% of the false positives in a state of the art nodule detection algorithm can be removed. The vessel-based volume registration algorithm uses both junctions and linear vessel segments as landmarks. Smoothness constraints of the transformation field are enforced by checking the curvatures of the transformation field. In the proposed registration algorithm, a novel way is proposed to use the constraints of linear vessel segments to improve the registration. The matches of linear vessel segments are found by actively deforming the vessel contours towards their matches on the target image. Compared with the commonly used registration algorithms, the proposed registration approach is more robust to noise and more accurate in estimating the actual transformation between temporal CT scans. By incorporating the vessel enhancement filter and the above mentioned constraints, the proposed registration approach is also better in handling the partial volume effects and the structural differences such as new nodules. |
