Fully elastic multimodality image registration using mutual information

We show that mutual information based registration over vanishingly small neighborhoods is equivalent to a particular form of gradient matching. Using the expression for the mutual information in terms of image gradients, we develop a voxel-wise registration algorithm that maximizes the sum of local mutual information between two images around all voxels. The similarity criterion does not require estimation of probability densities, and yields an analytical derivative suitable for gradient-based optimization methods. Simulation results indicate that the method achieves accurate nonlinear registration between different modality images from two different individuals.