Groupwise registration based on hierarchical image clustering and atlas synthesis

Groupwise registration has recently been proposed for simultaneous and consistent registration of all images in a group. Since many deformation parameters need to be optimized for each image under registration, the number of images that can be effectively handled by conventional groupwise registration methods is limited. Moreover, the robustness of registration is at stake due to significant intersubject variability. To overcome these problems, we present a groupwise registration framework, which is based on a hierarchical image clustering and atlas synthesis strategy. The basic idea is to decompose a large-scale groupwise registration problem into a series of small-scale problems, each of which is relatively easy to solve using a general computer. In particular, we employ a method called affinity propagation, which is designed for fast and robust clustering, to hierarchically cluster images into a pyramid of classes. Intraclass registration is then performed to register all images within individual classes, resulting in a representative center image for each class. These center images of different classes are further registered, from the bottom to the top in the pyramid. Once the registration reaches the summit of the pyramid, a single center image, or an atlas, is synthesized. Utilizing this strategy, we can efficiently and effectively register a large image group, construct their atlas, and, at the same time, establish shape correspondences between each image and the atlas. We have evaluated our framework using real and simulated data, and the results indicate that our framework achieves better robustness and registration accuracy compared to conventional methods.