Improved inter-modality image registration using normalized mutual information with coarse-binned histograms

In this paper we extend the method of inter-modality image registration using the maximization of normalized mutual information (NMI) for the registration of [18F]-2-fluoro-deoxy-D-glucose (FDG)-positron emission tomography (PET) with T1-weighted magnetic resonance (MR) volumes. We investigate the impact on the NMI maximization with respect to using coarse-to-fine grained B-spline bases and to the number of bins required for the voxel intensity histograms of each volume. Our results demonstrate that the efficiency and accuracy of elastic, as well as rigid body, registration is improved both through the use of a reduced number of bins in the PET and MR histograms, and of a limited coarse-to-fine grain interpolation of the volume data. To determine the appropriate number of bins prior to registration, we consider the NMI between the two volumes, the mutual information content of the two volumes, as a function of the binning of each volume. Simulated data sets are used for validation and the registration improves that obtained with a standard approach based on the Statistical Parametric Mapping software.