Rigid, multi-rigid, and non-rigid image registration of skeletal structures

The goal of my dissertation is to develop medical image registration methods for skeletal structures. It was motivated by the study of combining CT and MR images as a noninvasive alternative to CT-Myelography (CTM) for the diagnosis and surgical planning of degenerative cervical spine disease. While an individual bone can be modeled as a rigid body, the entire spine cannot because it is only piecewise rigid. This property of articulated movement makes the multi-rigid approach a plausible strategy: the entire spine on the CT images is segmented into a collection of disconnected vertebrae, each of which is then registered with the corresponding vertebra on the MR images. The key of the multi-rigid approach is thus the ability to segment each rigid body accurately for the subsequent registration. This is a challenging problem for the spine especially in patients with degenerative neck disease. To solve it, we developed a segmentation method called contour competition. This method represents each vertebra by a level set function, and all vertebrae evolve simultaneously to determine the boundary between them. Once segmentation is complete, rigid body registration is then used to estimate the transformation from the CT to the MR for individual vertebrae, and finally the CT images are superimposed on the MR scans to obtain the fusion images. The multi-rigid approach generated two useful results: the set of disarticulated bones and the set of transformation parameters for each of them. The former allow us to rapidly develop patient-specific models for the bones, and the latter enable us to study the movement of the bones from medical imaging. The multi-rigid approach has been applied to the analysis of ankle motion from MRI, for which we further improved the segmentation method by combining the graph cuts technique and the level sets method. Finally, we studied non-rigid image registration methods for template-subject matching problems. We used these methods to build statistical shape models to compare the morphology of the talus of different foot types, and to automatically determine local coordinate systems for the foot bones.