Assessing Spline-Based Multi-resolution 2D-3D Image Registration for Practical Use in Surgical Guidance

Abstract

A spline-based multi-resolution 2D-3D image registration algorithm has recently been introduced [1-3]. However, its accuracy, robustness, and efficiency have not been fully investigated. In this paper, we focus on assessing and improving this newly introduced 2D-3D registration algorithm. A phantom and a cadaver test, together with their respective ground truths, were specially designed for this purpose. A novel least-squares normalized pattern intensity (LSNPI) similarity measure was proposed to improve the accuracy and robustness. Several parameters that may also affect its robustness, accuracy, and efficiency are experimentally determined, including the final resolution level, the initial guess of the patient pose, the number of 2D projection images, and the angle between 2D projection images. Our experiments show that it is feasible for the assessed 2D-3D registration algorithm to achieve sub-millimeter accuracy in a realistic setup in less than two minutes, when it is used together with the newly proposed similarity measure. © Springer-Verlag 2004.