Sparsity-learning-based longitudinal MR image registration for early brain development

Longitudinal sequences of infant brain MR images are increasingly applied in early brain development studies, while their registration are highly challenging as rapid brain development causes drastic image appearance changes. To this end, we propose a novel sparsitylearning- based strategy to tackle the longitudinal registration of infant subject. First, we prepare a set of intermediate sequences, whose longitudinal (voxel-to-voxel) correspondences are established in advance. For each time point of the subject, we then utilize sparsity learning to identify its correspondences in the intermediate images at the same age and thus of similar appearances. Next, the intermediate sequences are used to bridge the temporal “gaps” between different subject time points, while the sparsity-learning-based correspondence detection is jointly conducted for all subject images to impose the temporal consistency. Finally, the deformation field of each subject time point is reconstructed from the spatio-temporal correspondences. Experimental results show that our method is able to achieve the longitudinal registration of the infant subject despite its varying appearances along time.