Affine registration: a comparison of several programs

Several registration programs with an affine model for the displacement field were tested on various 2D and 3D MRI of the same modality. The following programs were considered: AIR 3.0 (Woods, J. Comp. Assist. Tomogr, 22(1): 139–152, 1998), COCGV (Ostuni, JMRI, 7(2): 410–415, 1997), FLIRT (Jenkinson, Med. Image Analysis, 5(2): 143–156, 2001), Intramodal Registration (Thevenaz, IEEE Trans. Image Proc., 7(1): 27–41, 1998), SPM (Friston, Human Brain Mapping, 2: 165–189, 1995), and Patch Algorithm (Zhilkin, MRI 18(9): 1143–1150, 2000). Although some of these programs can perform multimodal registration, none was used in such a mode. This paper attempts a fair comparison of the performance of the Patch Algorithm with other programs. However, different settings of the programs' parameters may further improve the quality of the registration and/or change execution speed. The registered images, the CPU time required to perform the registration, and the error between the registered and reference images, are presented. Most of the programs give comparable accuracies of registration, but their execution times vary considerably. In general the AIR and Patch Algorithm require the least time. The Patch Algorithm can be easily parallelizable on a multi-processor computer.