Mutual information matching in multiresolution contexts

Image registration methods based on maximization of mutual information have shown promising results for matching of 3D multimodal brain images. This paper discusses the effects of multiresolution approaches to rigid registration based on mutual information, aiming for an acceleration of the matching process while maintaining the accuracy and robustness of the method. Both standard mutual information and a normalized version are considered. The behaviour of mutual information matching in a multiresolution scheme is examined for pairs of high resolution magnetic resonance (MR) and computed tomography (CT) images and for low resolution MR images paired with either positron emission tomography (PET) images or low resolution CT images. Two methods of downscaling the images are compared: equidistant sampling and Gaussian blurring followed by equidistant sampling. The experiments show that a multiresolution approach to mutual information matching is an appropriate method for images of high (sampling) resolution, achieving an average acceleration of a factor of almost 2. For images of lower resolution the multiresolution method is not recommended. The little difference observed between matching with standard or normalized mutual information seems to indicate a preference for the normalized measure. Gaussian blurring of the images before registration does not improve the performance of the multiresolution method