Abstract

MRI slice stacking involves retrospective combination of 2D MRI images to form pseudo 3D volumes. It is useful because physical constraints limit the temporal/spatial resolutions with which dynamic 3D MRI volumes can be acquired and so stacking fast high-resolution 2D images can yield pseudo 3D volumes with high in-plane spatial and temporal resolution. However, it is important that the stacked 2D images were acquired at consistent motion states. Assessing motion state consistency between slices representing different anatomy is challenging as the image contents are not easily comparable. Manifold alignment (MA) is a technique which provides a solution to this problem by embedding the 2D images for all slices into one globally consistent low-dimensional space. One successful approach to MA involves forming graphs from each slice dataset and using graph descriptors to find correspondences between datasets. Here we propose a new graph descriptor for the slice stacking problem, inspired by work in the computer vision literature, and evaluate it with two experiments. First, using a highly realistic synthetic MRI dataset in which reconstructed volumes can be compared to a ground truth, we find our method significantly outperforms the state of the art. Second, we use in vivo MRI data and show that the volumes reconstructed by our method have a higher degree of self-consistency.