Medical Image registration with B-Spline Transformation
- Medical image registration is an important task in various applications such as temporal analysis of medical images from a single modality (where structural or anatomical changes over time are studied), multi-modal image fusion (in which data acquired from different imaging devices or protocols is fused to help with diagnosis and treatment planning), or atlas matching (where uses normal anatomical variability). Establishing functional and/or spatial anatomical correspondences between different image acquisitions is the goal of image registration. As rigid alignment might be insufficient to describe variations of images in certain applications where deformations play a role, we should rely on non-rigid image registration. Non-rigid registration methods are capable of aligning images where correspondence cannot be achieved without localized deformations and can therefore better accommodate anatomical, physiological, and pathological variability between patients. Elastix image registration software is employed in this thesis to conduct our experiments. In this thesis, we use interpolation-based B-spline transformation models that parameterize a Free-Form Deformation Field. B-splines interpolate the dense deformation field from a given set of control points. B-spline transformation model assumes the smoothness and continuity of the deformation field on the image. However, when aligning two images that contain sliding contact, B-spline transformation leads to an inaccurate registration result, and the deformation field is not always smooth or the deformation field is always discontinuous, but not necessarily everywhere. When the knee is bent, the sliding contact occurs at the knee cartilage and around the capsular tissues. Different types of discontinuities in the overall deformation field may occur when sliding contact occurs. Imposing smoothness and continuity in the registration will generate artifacts in the consequent deformations, as well as any quantities derived from it. To handle the artifacts in the resulting deformations, one can use a regularizer or implement the sliding motion in the transformation model. Hence, in this thesis, we compare the B-spline registration, B-spline with Transform Rigidity Penalty, B-spline with Distance Rigidity Penalty, and the sliding motion with B-splines. B-spline with Transform Rigidity Penalty (TRP) and B-spline with Distance Rigidity Penalty (DPRP) models use a regularizer to handle the sliding motion. However, the sliding Motion with B-splines model handles sliding motion in the transformation model itself. The chosen registration parameters and registration algorithm were assessed first using three synthetic datasets. In a second validation, Knee MRI images that belong to two different patient datasets are registered on a temporal sequence of each patient. Registration accuracy was assessed based on the corresponding segmented bones and neighboring soft tissue. Compared to B-spline approach, the B-spline with TRP, B-spline with DPRP, and the sliding motion with B-splines gave deformation fields that represented more plausible image correspondences, while giving rise to comparable registration accuracy. In one dataset, the B-spline with TRP and B-spline with DPRP models reflected the deformation better than other approaches. However, sliding motion with B-splines model outperformed all other approaches in the other dataset.
| Author: | Mahmut Özmen |
|---|---|
| Document Type: | Master's Thesis |
| Granting Institution: | Freie Universität Berlin |
| Advisor: | Alexander Tack, Stefan Zachow |
| Date of final exam: | 2022/02/10 |
| Year of first publication: | 2022 |
| Page Number: | 64 |

