Non-rigid Multimodal Medical Image Registration Based on Local Variability

Abstract

In this work, we present a novel approach for multimodal elastic registration of medical images, where the key idea is to use local variability measures based on entropy, variance or a combination of these metrics. The proposed methodology relies on finding the displacements vector field between pixels of a source image and a target one, using the following three steps: first, an initial approximation of the vector field is achieved by using a parametric registration based on particle filtering between the images to align; second, the images previously registered are mapped to a common space where their intensities can be compared; and third, we obtain the optical flow between the images in this new space. To evaluate the proposed algorithm, a set of computed tomography and magnetic resonance images obtained from different views, were modified with synthetic deformation fields. The results obtained with the four proposed local variability measures show an average error of less than 1.4 mm, and in the case of the entropy less than 1 mm. In addition, the convergence of the algorithm is highlighted by the joint entropy. Therefore, the described methodology could be considered as a new alternative for multimodal elastic registration of medical images.