Multi-modality image registration with gradient orientation information based on entropic spanning graph

Recently a growing interest has been seen in multi-modality image registration with graph based entropy estimator. As contrasted to the plug-in estimator, entropic spanning graph estimates the entropy directly, without the need to compute probability density function. Graph based entropy estimator usually works directly with pixel intensity. However, pixel intensity is sensitive to intensity change introduced by multi-modality images. In this paper, a novel feature is proposed to capture the structural information based on entropic spanning graph, which remains unchanged between the multi-modality images. We construct the new feature for each sample point through combination of gradient orientation and magnitude, and considering gradient orientation in the interval [0°, 180°) instead of [0°, 360°). This makes the new feature invariant to contrast reversals, translation and rotation. Experiment results show that the algorithm is robust to nonlinear intensity variation and achieves better success rate while maintaining good registration accuracy.