Purpose: Pulmonary vascular trees can be automatically extracted from lung CT images using established methods. The purpose of this study is to test the hypothesis that the extracted vascular trees can be used to improve deformable image registration (DIR) accuracy for lung CT images. If it is approved, the proposed method provides a generic yet practical way to improve lung CT DIR accuracy for any DIR algorithm, and the DIR results can then be quantitatively measured using tree alignment and visually verifiable by human observers.
Methods: A 3D vascular tree probability map algorithm was optimized to detect pulmonary vascular trees in lung CTs. Demons DIR algorithm was then applied to register 4DCT image pairs using extracted vascular trees. DIR performance was evaluated on 10 benchmark datasets from DIRLAB. Target registration errors (TREs) on ground-truth landmark pairs were computed and compared to those generated by the state-of-the-art DIR algorithms.
Results: By registering vascular trees instead of CT images, Demons performed almost as well as state-of-the-art DIR algorithms in both accuracy and robustness. Average 3D TREs by Demons over 10 cases (1.06Â±1.15 mm) were comparable to the best reported TREs (0.91Â±1.07 mm, www.dir-lab.com/results.html) For Case #2, Demons even outperformed the best reported algorithm. In addition, average TREs by isoPTV, the best DIR algorithm for lung 4DCT, were 0.95Â±1.15 mm, and Demons outperformed isoPTV in 3 of 10 cases.
Conclusion: Pulmonary vascular trees are relatively easy to extract from lung CT images and proved to be useful for improving DIR accuracy and robustness. Although only demonstrated using the aged Demons algorithm, the proposed procedure is generic and applicable to any DIR algorithms.
Funding Support, Disclosures, and Conflict of Interest: Supported by AHRQ R01HS022888