Cross-manifold Guidance In Deformable Registration Of Brain Mr Images

Deformable image registration is very important for many brain studies based on MR images.Traditionally,image registration is regarded as an optimization problem,which tries to find a deformation field to maximize the similarity between the warped subject image and the template image.However,the optimization is often cursed by the high dimensionality of the images and deformation fields,making it difficult to solve when the images have large appearance differences.Recent studies have found that,within an image collection,a manifold can be used to characterize the high-dimensional distribution of a set of brain MR images.Thus,the deformation field between two images can be seen as a geodesic pathway on the complex manifold.This results in the image guidance in pairwise registration.That is,if the subject image and the template image have very different appearances,which means they don't distribute closely on the manifold,the subject image can utilize other intermediate images on the manifold as the bridges to fill in the gaps along its deformation pathway towards the template image.But to represent the manifold and provide guidance,a lot of other images have to be already registered to the template image.Besides,it is still non-trivial to estimate the deformation pathway directly because of the intrinsic complexity of the manifold.In this dissertation,we break the restriction of utilizing one single and complex manifold to provide guidance for pairwise registration,by short-circuiting the subject-template pathway with routes from multiple simpler manifolds.Specifically,we reduce anatomical complexities of the subject and template images to project them to virtual and simplified manifolds.The projected simple images then guide the subject and template image to simpler manifolds where the subject and the template are much more similar in appearance and their registration can be done more easily and accurately.Finally,the subject-template pathway,is attained by traversing multiple manifolds of lower complexity rather than computing on the original single complex manifold.We apply proposed cross-manifold guidance to brain MR image registration.Our experiments show that our method leads to superior alignment accuracy compared to state-of-the-art deformable registration techniques under the same situation and cross-manifold guidance can be a very useful scheme for refining image registration.