Automatic Landmark Detection and Norid Registration of Intra-Subject Lung CT Images

In order to improve the accuracy of lung registration, a new hybrid non-rigid approach is proposed to deform the intra-subject lung images with the assistance of the anatomical landmark information. Firstly, the two volumes, respectively for expiration and inspiration breath-hold lung, are segmented to extract the airways which have tree-like topology. Secondly, the two extracted airways are skeletonized by an efficient voxel-coding based method. Thirdly, the points with distinct topological features in the two breath-hold skeletons are picked as the landmarks and the correspondence anatomical relationship between the landmarks are described by Huffman-coding. Finally, a coarse-to-fine registration scheme, combined with landmark-based algorithm (thin-plate spline, TPS) and intensity-based (modified Demons) algorithm is proposed to warp and deform the intra-subject volumetric CT images. The experimental results show that the proposed approach is more accurate and inverse-consistent owing to the intrinsic one-to-one mapping of correspondent points and improved registration strategy.