An Improved Multi-Stage Method for Medical Image Registration Based on Mutual Information

The objective of registration process is to obtain a spatial transformation of a floating image to a reference image by which a similarity measure is optimized between the two images. A widely used measure is Mutual Information (MI). Registration based on mutual information is robust and dataindependent and could be used for a large class of monomodality and multimodality images. This method requires estimating joint histogram of the two images. As a result, it requires an extremely high computation time. This is its main drawback especially when it is applied to volume images. In order to speed up the registration process, a multi-resolution approach has been introduced before. In this method, a pyramid of low to high resolution images is used to find a rough estimate of parameters of the optimum transformation in a relatively short time using low resolution images, which is subsequently used as initial value for the higher resolutions. An appropriate estimation of the transformation parameters improves speed of the optimization algorithm. In this paper, we present a new improved method of sample selection for multi-stage registration based on mutual information. Instead of downsampling technique used in the pyramid methods, we propose a new technique to find a suitable subset of image samples, which results in a better estimate of the optimal transformation. A comparison for MR images indicates that our proposed method yields a better registration than subsampling method, especially when subsampling factor is low. Moreover, the experimental results involving three-dimensional clinical images of CT, MR and PET are presented for rigid registration.