Sparse Patch-Guided Deformation Estimation for Improved Image Registration

Abstract

This paper presents a novel patch-guided initial deformation estimation framework for improving performance of the existing registration algorithms. It is challenging for the registration algorithm to directly align two images with large anatomical shape difference, when no good initial deformation is provided. Inspired by the patch-based multi-atlases segmentation method, we propose to estimate the initial deformation between two images (under registration) in a patch-by-patch fashion. Specifically, after obtaining the sparse representation for each local patch in the subject by using the training patches in the over-complete dictionary that include both patches and their associated deformations from the training images, the initial deformation for each local subject patch can be predicted by those estimated sparse coefficients. More specifically, after registering all training images to the template in the training stage, the following three steps can be used to register any given subject image. First, for each key point in the subject, we can construct a coupled dictionary from the nearby patches in the training images and their associated deformations, and can then use this dictionary to seek for sparse representation of the respective subject patch. The estimated sparse coefficients can be used to fuse the associated deformations in the dictionary, for estimating the initial deformation for the respective subject key point. Second, after estimating the initial deformations on a small number of key points in the subject, thin-plate spline (TPS) can be applied to interpolating the dense deformation field. Finally, we can apply any existing deformable registration method (with reasonable performance) to estimate the remaining deformation from the template to subject. Experimental results on both simulated and real data show that our patch-guided deformation estimation framework can allow for more accurate registration than the direct use of the original methods for registration.