Textural mutual information based on cluster trees for multimodal deformable registration

Mutual information (MI) has been widely used in image analysis tasks such as feature selection and image registration. In particular, it is the most widely used similarity measure for intensity based registration of multimodal images. However, a major drawback of MI is that it does not take the spatial neighbourhood into account. An effective way of incorporating spatial information could be of great benefit in a number of challenging applications. We propose the use of cluster trees to efficiently incorporate textural information from the local neighbourhood of a voxel into the computation of MI, while at the same time limiting the number of bins used to represent this higher-order information. This new similarity metric is optimised using a Markov random field (MRF). We apply our new method to the registration of dynamic lung CT volumes with simulated contrast. Experimental results show the advantages of this technique compared to standard mutual information.