Nonrigid Registration Of Medical Images Based On Anisotropic Structure Tensor And Joint Saliency Map

Nonrigid registration has been applied in many areas such as computer vision, remote sensing, topographical reconnaissance, pattern recognition and medicine since it is able to describe the process of deformation in a more realistic way. However, there are still lots of difficulties in dealing with this kind of registration. In medical registration, for example, it remains a challenge in coping with the alignment of pre- and intra-operative images after tumor resection because tumor resection introduces missing correspondence between images as well as local large deformation with the motion of other tissue. In this thesis, we proposed a novel method using anisotropic local structure and joint saliency map to solve both missing correspondence and local large deformation.The algorithm is based on the framework of pyramidal block matching scheme combined with mixed similarity measure strategy which means we choose cross correlation for low resolution case while mutual information for high resolution case. The spatial regularization of discrete displacement field after block matching is achieved by normalized convolution (NC). To avoid the irregularity of deformation field resulting from the traditional NC, we focus on the improvement of two key factors- applicability function and data certainty.To remedy the limitation of isotropic applicability function, we designed an anisotropic Gaussian kernel as an applicability function whose shape and orientation are adaptive to the local structure tensor (LST). Thus, the main axis of Gaussian kernel is aligned with the eigenvector of LST and the scale in each direction is related to the eigenvalues. Therefore, based on the anisotropic applicability function, the localization of the convolution operator is defined according to the local image structure. Additionally, we introduce the concept of joint saliency map, first presented in our previous work, into deformation regularization. The data certainty of NC is assigned based on the joint saliency value calculated from the similarity measure of two LSTs of the corresponding points. In accordance with the joint saliency map, high certainty is given to the pixels(voxels) belonging to joint salient regions while a extremely low certainty is assigned to those from background or isotropic regions or outliers, the registration will concentrate on joint salient regions in order to repress the negative impact of outliers and isotropic regions.The experimental results show that our approach is effective to avoid the irregularity as well as discontinuities of deformation field and secures a smooth and continuous deformation field, and also the registration maintains the local image structure to some extent. Moreover, owing to the consideration of joint saliency between images, our method is robust to outliers and is sufficiently accuracy compared with some other state-of-the-art algorithms.