Anatomical Markov prior-based multimodality image registration

Abstract

Some similarity measures used in state-of-the-art multimodality image registration algorithms, (e.g., mutual information (MI)) have been shown to be suitable anatomical priors for maximum a posteriori reconstruction in emission tomography. Therefore, it is reasonable to assume that some originally designed anatomical priors may also be well suited for multimodality image registration. In this work, we evaluate the registration performance of three variants of an anatomical Markov prior, previously proposed by Bowsher et al. First, simulated data are used to verify whether the suggested registration criteria yield an optimum when an FDG positron emission tomography (PET) image and a T1-weighted magnetic resonance (MR) image of a human brain are perfectly aligned. Next, the registration accuracy of the proposed criteria is assessed for PET to MR and MR to PET registration of simulated human brain images, and compared to the accuracy reached by MI. Last, the new methods are applied to challenging measured rat and mouse brain data sets, consisting of low resolution FDG microPET images and high resolution microMR images with a strong bias field. It was shown that the anatomy-based Markov priors indeed yield a well-defined optimum for aligned PET-MR images and that similar registration accuracy can be achieved as with MI, especially for registration to MR images suffering from a bias field. Nevertheless, in contrast to MI, the new criteria usually require a good initial guess of the transformation parameters in order not to get stuck in a local optimum. The proposed methods are shown to be superior to MI for registering measured microMR brain images with a strong bias field to FDG microPET images if a good initialization is provided.