Registration by maximization of mutual information-a cross validation study

Abstract

Mutual Information (MI), or relative entropy has been used as a similarity criterion in medical image registration. MI is a measure of the dispersive behavior of the joint histogram of geometrically related voxels' intensities in both images. This dispersion is assumed to be smaller when the images are aligned. Besides, no assumptions are made, before bringing images together or during MI calculation, regarding the nature of the relation between corresponding voxels. In this work the authors assess how the elaboration of joint histogram influences overall accuracy of maximization of MI registration method in unimodality and multimodality registration. For this purpose, across validation study is performed considering two other widespread registration algorithms: the SPM's registration package and multimodality AIR method. The correct elaboration of the joint histogram depends not only in the interpolation function used to resample the test image, but also in the re-scaling procedure used to fit images' values in the joint histogram and its subsequent update. The authors also evaluate if histogram's dimensions are important for overall accuracy.