DIS-CO Lung : Fast and Accurate Lung CT Alignment Using Discrete Keypoint Detection and Continuous Image Registration

Despite great research effort over the previous years, accurate registration of thoracic CT scans remains challenging. In particular, large motion between full inhalation and exhalation leads to unacceptably high alignment errors. We address these challenges by combining discrete and continuous optimization and have developed a new algorithm which performs accurate sparse-to-dense alignment of thoracic CT scans. First, a discrete keypoint detection and matching using sparse graph-based correspondences is performed. Second, a continuous, deformable image registration incorporating dense intensity and the aforementioned correspondence information is employed. A volume change control mechanism is used that limits local volume change and prohibits foldings of the deformation grid. The method achieves high accuracy and very competitive run times of less than 5 minutes per registration.