A New Image Similarity Measure with Reduced Sensitivity to Interpolation and Generalizability to Multispectral Image Registration

Abstract

Mutual information (MI) has proven to be a useful similarity measure for spatial registration between related pairs of images in various medical imaging applications. Image registration algorithms that utilize the MI assume that the best alignment between a pair of images is reached when their MI is at its maximum. However, this assumption is not always valid because the MI is not only sensitive to dissimilarity between images, but also to the image interpolation operations performed during the optimization process in image registration algorithms. When the images that are being registered are close to their optimum spatial alignment, MI's sensitivity to interpolation may become dominant over its sensitivity to image misalignment, hence limiting the accuracy of the image registration method. In this paper, we present an entropy-based cost function, closely related to MI, that can be made relatively insensitive to interpolation effects, and can be generalized to registration of multispectral images.