Image registration based on dynamic directed graphs with groupwise image similarity

We propose a groupwise image registration method using sparse coding and graph theoretic techniques. A sparse coding method is used to estimate image similarity measures among images to registered, yielding asymmetric, groupwise image similarity measures for each image to other images in the group. Based on the asymmetric groupwise image similarity measures among different images, a directed graph is built for learning a manifold of images so that a group center image can be identified and each image can be registered to the center image following a directed shortest path which decomposes a large deformation between two images into a series of small and anatomically meaningful deformations. The groupwise image similarity measures are progressively updated, so does the directed graph, facilitating accurate alignment of a group of images following directed shortest paths derived from the dynamic directed graphs. Compared with state of the art graph based registration methods, our method is more robust and able to achieve better registration accuracy.