Robust methods for human airway-tree segmentation and anatomical-tree matching

Robust and accurate segmentation of the human airway tree from multi-detector computed-tomography (MDCT) chest scans is vital for many pulmonary-imaging applications. Accurately segmented airways are particularly important for image-based bronchoscopic guidance. Here, the segmented airways provide a patient-specific 3D airway model, which is used to generate Virtual Bronchoscopic (VB) images simulating the view of a real bronchoscope inside the patient's airways. The VB views are compared with the live bronchoscopic video to lead the bronchoscopist to a predetermined airway site, where a diagnostic tissue sample is taken. VB guidance has a demonstrated potential for improving the diagnostic utility of bronchoscopy, especially for peripheral sites located many generations deep into the airway tree. Existing segmentation methods are insufficient for peripheral VB guidance, however, as they frequently fail to segment small peripheral airways with weak image signatures.;This thesis also addresses the problem of matching pairs of anatomical trees depicted in two different high-resolution 3D images. Three basic steps are used to match the trees: (1) image segmentation, to extract the raw trees from the 3D image data; (2) axial-analysis, to define the underlying centerline structure of the trees; and (3) tree matching, to determine corresponding branches and branchpoints between the centerline structures of the trees. This thesis focuses on step (3). The matching task is complicated by several problems associated with current segmentation and axial-analysis methods, including missing branches, false branches, and other topological errors in the extracted trees. A model-based approach is proposed in which the input trees are assumed to arise from an initially unknown common tree corrupted by a sequence of topological deformations. A set of valid matches consistent with this model is constructed and the optimal match is defined to be the valid match that maximizes a similarity measure. We derive and analyze a dynamic programming (DP) algorithm that efficiently locates a globally-optimal match. The proposed algorithm is validated by comparing automatically-generated matches with ground-truth hand-generated matches. In particular, the proposed algorithm is shown to increase an important measure of matching accuracy from the 57% achieved by the best existing method to more than 96% for human airway-tree data. Finally, the proposed tree-matching methodology is shown to be useful in automatically labeling human airway-tree anatomy.;This thesis proposes a novel automatic airway-segmentation algorithm, which searches the entire lung volume for airway branch signals and poses segmentation as a global graph-theoretic optimization problem. The algorithm is extensively validated on scans of both healthy and diseased airways obtained from several different sources. In particular, comparisons with ground-truth airway segmentations demonstrate that the proposed algorithm extracts substantially more peripheral airways than existing methods while producing very few false-positive branches. The automatic algorithm is combined with a suite of interactive tools for cleaning and extending critical local areas of the airway tree to form a complete computer-based segmentation system. We present results from a clinical pilot study validating the utility of the full segmentation system. During this study, airway-tree segmentations produced by the proposed system were used to help plan and guide live peripheral bronchoscopic surgery in humans. The successful outcome of the pilot study was due, in large part, to the representational accuracy of patient-specific 3D airway models derived from segmented airways produced by the proposed system.