Airway Segmentation, Skeletonization, and Tree Matching to Improve Registration of 3D CT Images with Large Opacities in the Lungs

In this work, we address the registration of pulmonary images, representing the same subject, with large opaque regions within the lungs, and with possibly large displacements. We propose a hybrid method combining alignment based on gray levels and landmarks within the same cost function. The landmarks are nodes of the airway tree obtained by specially developed segmentation and skeletonization algorithms. The former uses the random walker approach, whereas the latter exploits the minimum spanning tree constructed by the Dijkstra’s algorithm, in order to detect end-points and bifurcations. Airway trees from different images are matched by a modified best-first-search algorithm with a specially designed distance function. The proposed method was evaluated on computed-tomography images of subjects with acute respiratory distress syndrome, acquired at significantly different mechanical ventilation conditions. It achieved better results than registration based only on gray levels, but also better than hybrid registration using a standard airway-segmentation method.