| A fundamental problem in medical image analysis is the integration of information from multiple images of the same subject, acquired using the same or different imaging modalities and possibly at different time. An essential aspect of this problem is image registration, which is the task of finding geometric relationships between corresponding points in multiple images of the same scene. Various registration methods have been proposed over recent years, among which registration strategies based on maximization of mutual information have been proved to be promising methods and have been widely used in multi-modality image registration.However, applying mutual information (MI) to original intensities only takes statistical information into consideration, while spatial information is completely neglected. In the first part of this dissertation, a novel approach is proposed to incorporate spatial information into MI through gradient vector flow (GVF). With this approach, MI now is calculated from the GVF-intensity (GVFI) map of the original images instead of their intensity values. The algorithm is implemented and applied to multi-modality brain image registration to test the accuracy and robustness of the proposed method. Experimental results show that the success rate of our method is higher than that of traditional MI-based registration.In many applications, a rigid transformation is insufficient to describe the spatial relationship between two images. Thus, elastic transformations, or non-rigid transformations are often required in image registration. In the second part of this dissertation, we present a generalized gradient-guided non-rigid registration strategy. The derivation procedure is similar to that by Lucas and Kanade, but in a more general manner. In experiments, we compare the proposed method and other gradient-guided methods in the literature, using both synthetic and real images. It is shown that methods combining gradients from both source and target images usually perform better.In the third part, we apply previously described registration methods to atlas-based brain magnetic resonance (MR) image segmentation. A pre-labelled image or atlas is first registered to the subject image to be segmented, and the deformation field for each voxel is derived. Then the structures delineated in the atlas are projected onto the subject image by applying the deformation field to the atlas mask. We validate our results using the datasets from IBSR. Quantitative comparisons using various criteria show that the proposed method is better than or comparable to published methods. |
