Multi-attribute combined mutual information (MACMI): An image registration framework for leveraging multiple data channels

We present a novel methodological framework for leveraging multiple image sources, including different modalities, acquisition protocols or image features, in the registration of more than two images via information theoretic data fusion. The technique, referred to as multi-attribute combined mutual information (MACMI), adopts a multivariate application of mutual information (MI) to allow several coregistered images to be represented as a single high dimensional multi-attribute image. Our approach improves scenarios involving registration of multiple images as it, (1) utilizes all aligned images obtained in earlier registration steps, (2) improves alignment accuracy compared with pairwise approaches that only consider two images (and hence a fraction of the available data) at a time, and (3) avoids complex optimization problems often associated with fully-groupwise methods. For example, if two coregistered volumes such as T2-weighted and PD-weighted MRI are to be aligned with PET, it is intuitively better to use information from both MR protocols instead of choosing one for registration with PET. In the automated elastic registration of 20 corresponding multiprotocol (T1, T2, PD) synthetic MRI images of the brain with known misalignment of PD MRI, MACMI showed significant improvement in terms of deformation field error over conventional MI-based pairwise registration (p <; 0.05). For a total of 108 corresponding whole-mount histology (WMH), T2 MRI, and DCE (T1) MRI images obtained from 17 prostate specimens with cancer, elastic registration of WMH to bothMRI protocols simultaneously was performed viaMACMI. Improved alignment in terms of prostate overlap and cancer localization was observed using MACMI, compared to pairwise registration of WMH to the individual T2 and DCE MR protocols.