Knowledge Based Image Segmentation Using Deformable Registration:Application To Brain MRI Images

Image segmentation plays an important role in medical image analysis. Medical image segmentation aims to divide an entire image into different non-overlapping regions. These regions represent different anatomical structures clearly. Segmenting and analyzing anatomical structures from brain MRI images is an effective method to diagnose brain diseases and track their evolvement. However the complexities of medical images make medical image segmentation a challenging task. Fortunately, the a priori knowledge provided by the atlas can contribute much to a better segmentation. Image registration is a good technique to fuse medical priori knowledge into segmentation procedure, which forms a new image segmentation technique, registration based segmentation. In registration based segmentation, an optimal spatial transformation deforming the source image into the target image is first obtained. Then the labeled structures in the atlas (correspondence to the source image) are mapped to the target image based on the obtained optimal spatial transformation. Finally, the required segmentation will be acquired easily. Obviously, the segmentation quality mainly depends on the registration quality.Using simple rigid registration to describe subtle local differences between individuals is insufficient due to the complexity of the human body structures. Therefore non-rigid registration techniques should be used to solve the problem. Non-rigid registration approaches usually have higher degree of freedom, and can describe more complex nonlinear deformations. Compared to the rigid registration, non-rigid registration is not mature. There are many problems to be solved, such as establishing a reasonable deformation model, improving the registration speed and accuracy, developing convincing algorithm evaluation criteria, and so on.Non-rigid registration based on physical models is an important branch in the current techniques, because they have unique abilities to describe the physical behaviors of organism. In this dissertation, we studied diffusion model based Demons non-rigid registration algorithm in depth, which originated from thermodynamic theory. The main researches focus on the intensity and spatial normalization, regularization method of the deformation field, the contributions of a priori knowledge to non-rigid registration and image segmentation. In segmenting deep brain internal structures applications, a hybrid intensity and shape knowledge non-rigid registration algorithm is proposed. The major contributions of this dissertation are as follows:1. Analyze, optimize, and implement diffusion model based Demons non-rigid registration algorithm.Demons non-rigid registration is one of excellent algorithms in physical model based registration. It is suitable to register two images within the same modality. In this dissertation, the Demons algorithm and its variant, active Demons algorithm, are studied in depth. Two important parameters, elasticity parameter and equalization parameter, and the limitations of the deformation force are analyzed carefully. Then instructive parameter setting principles and an improved deformation force are proposed.2. Propose a more localized non-rigid registration algorithm, which can complete spatial and intensity normalization simultaneously.Non-rigid registration algorithm using the sum of squared intensity differences as the similarity measure depends the image intensity knowledge absolutely. The assumption is that the intensities of two corresponding voxels are equal. However this condition is seldom fulfilled in real-world medical image registration without intensity normalization, because there are many factors that may affect observed intensities of a tissue over the imaged field, such as the different scanner or scanning parameters, normal aging, different subjects, and so on. In addition, there are large differences between individuals, which will influence the quality of inter-subject registration. Spatial normalization should also be done in advance. Combined with Demons algorithm, a step-wise non-rigid registration strategy is proposed. The proposed method not only can perform intensity and spatial normalization simultaneously, but also has higher localization ability.3. Propose a topology preserved Demons non-rigid registration algorithm.One disadvantage of Demons algorithm is that the topological invariance can not be ensured in theory, though bijectivity and deformation field smoothing technologies are adopted. The obtained deformation field is a dense vector field, which has a high deformation freedom. If there are no constraints to the deformation field, the spatial transformation can not be guaranteed reasonable and realizable. Through analyzing the critical points of a vector field, a topology preserved Demons non-rigid registration algorithm is proposed. In the context of preserving the characteristics as much as possible, the original deformation field changes to be topology preserved.4. Propose a hybrid intensity and shape features non-rigid registration algorithm.The deep brain internal structures, such as the caudate nucleus, putamen, thalamus, hippocampus, have close relationships with Parkinson's syndrome, epilepsy, dementia, Creutzfeldt-Jakob and other brain diseases. Analyzing the size and shape changes will contribute to the clinical diagnosis. However, these structures have complex shapes, small sizes, fuzzy boundaries and large partial volume effect in MRI images. Therefore, segmenting deep brain internal structures is a challenging task. Among different segmentation methods, registration based segmentation method is promising. By analyzing basic principles and the weakness of the intensity based non-rigid registration algorithm, a new hybrid intensity and shape features non-rigid registration algorithm is proposed. It is suitable for multi-object segmentation problem. Better results are obtained when the proposed algorithm is used to segment deep brain internal structures.