Abstract

Most existing methods for registration of three-dimensional tomographic images to two-dimensional projection images use simulated projection images and either intensity-based or feature-based image similarity measures. This paper suggests a novel class of similarity measures based on probabilities. We compute intensity distributions along simulated rays through the 3-D image rather than ray sums. Using a finite state machine, we eliminate background voxels from the 3-D image while preserving voxels from air filled cavities and other low-intensity regions that are part of the imaged object (e.g., bone in MRI). The resulting tissue distributions along all rays are compared to the corresponding pixel intensities of the real projection image by means of a probabilistic extension of histogram-based similarity measures such as (normalized) mutual information. Because our method does not compute ray sums, its application, unlike DRR-based methods, is not limited to X-ray CT images. In the present paper, we show the ability of our similarity measure to successfully identify the correct position of an MR image with respect to a set of orthogonal DRRs computed from a co-registered CT image. In an initial evaluation, we demonstrate that the capture range of our similarity measure is approximately 40 mm with an accuracy of approximately 4 mm.

Citations