Intensity-based two-dimensional-three-dimensional medical image registration
Carlo Tomasi, Daniel B. Russakoff · 2004
Intensity-based 2D-3D medical image registration is a special case of the pose estimation problem from computer vision with many applications in medicine. The task is to determine the pose of a preoperative CT (3D) image using one or more intraoperative X-ray projection (2D) images. We present an overview of the 2D-3D intensity-based image registration problem in the medical domain as well as results from several methods we have developed to aid in its practice. In particular, we present four results: (1) The generation of synthetic X-ray projection images, known as digitally reconstructed radiographs (DRRs), is typically the most computationally expensive step in intensity-based 2D-3D registration algorithms. We introduce attenuation fields, an extension of light field rendering techniques from the graphics community to generate DRRs several orders of magnitude faster than was previously possible using conventional methods. (2) We present a full 2D-3D registration algorithm using attenuation field DRRs and validate its accuracy using real, clinical data with known ground truth. We also use this algorithm and data set to compare the efficacy of several well-known similarity measures. (3) We present a new, hybrid similarity measure that is a weighted combination of an intensity-based image similarity measure and a point-based measure incorporating a single fiducial marker. (4) Finally, we discuss a novel similarity measure we have developed called regional mutual information (RMI). RMI is an extension of mutual information which incorporates spatial information in a principled way. The additional spatial information helps make its use as a similarity measure much more robust to initial misregistration than standard mutual information.