Groupwise Image Registration with Atlas of Multiple Resolutions Refined at Test Phase

Ziyi He, Tony C. W. Mok, Albert C. S. Chung

Abstract: Groupwise image registration (GIR) is a fundamental task that facilitates the simultaneous deformation of a group of subjects towards a specified or implicit center. Existing works mainly focus on either optimization-based methods that provide superb results but consume substantial time, or learning-based methods that are efficient but lack the flexibility to generalize across different domains and scales. To leverage the advantages of both methodologies, we present a robust method, Test-time Atlas adaptation for Groupwise registration (TAG), which generates a high-quality, group-specific atlas for groups of varying resolutions. Our method allows training at the test phase on target groups based on a learning-based GIR framework that bridges the gap between diverse groups. Besides the refinement of atlases at the original resolution, we propose additional modules to extend the scheme to groups of higher or lower resolutions at little cost. The method is evaluated on 3D brain MRI datasets to demonstrate its effectiveness. Evaluations of the registration accuracy and unbiasedness of atlases illustrate that TAG outperforms state-of-the-art benchmarks and maintains flexibility and robustness under a variety of scenarios.