Feature-constrained Nonlinear Registration of Lung CT Images

Deformable image registration is a key enabling technology for advanced treatment of lung cancer patients, as it can facilitate motion estimation, structure segmentation, as well as dose tracking and accumulation. In this work, we developed a hybrid feature-constrained deformable registration method and applied it to tackle the EMPIRE10 (Evaluation of Methods for Pulmonary Image Registration 2010) lung image registration challenge. The method uses the results of feature detection and matching based on robust 3D SURF descriptors to guide an intensity-based deformable image registration. The initial registration result is further refined by a hybrid MI/NSSD deformable registration process. The proposed method is fully automatic and does not require pre-segmentation of any lung structures. Validation results on the EMPIRE10 data showed that our method performed very well among 34 competing algorithms. Future improvement is possible with adaptive parameter selection, site-specific feature detection methods, and better deformation models.