An information theoretic approach for non-rigid image registration using voxel class probabilities

Abstract

We propose a multimodal free-form registration algorithm that matches voxel class labels rather than image intensities. Individual voxels are displaced such as to minimize the Kullback-Leibler distance between the actual and ideal joint probability distribution of voxel class labels, which are assigned to each image individually by a previous segmentation process. We evaluate the performance of the method for inter-subject brain registration with simulated deformations, using a viscous fluid model for regularization. The root mean square difference between recovered and ground truth deformations is smaller than 1 voxel.