Monomodal image registration using mutual information based methods

Image registration methods based on mutual information criteria, including mutual information and normalized mutual information, have been widely used in 3-D multimodal medical image registration and have shown promising results. Although they are also used in monomodal image registration, their performance is not as excellent as that in multimodal registration. There are many fluctuations in the registration function, which hinder the optimization procedure and lead to registration failure. This paper discusses this problem and ascribes it to interpolation artefacts and the variability of entropy. We implement experiments to evaluate the performance of the two similarity measures for 2-D and 3-D monomodal registration. To avoid the interpolation artefacts, we use pixels or voxels as the translation metric; to diminish the influence of entropy variability, we use normalized mutual information. The results show that, both for standard and normalized mutual information, the fluctuations caused by interpolation are fewer in the function of the registration without interpolation. Normalized mutual information has some similar properties to mutual information, but is almost invariant to the changing of entropy and appears to be more stable and robust than standard mutual information. These differences seem to indicate a preference for the normalized mutual information in monomodal registration.