2D-3D Rigid-Body Registration

The Problem
Registration of two datasets is the process of identifying a geometrical transformation that locates the coordinate system of one in that of the other. Rigid body registration restricts the searched transformation to be a combination of translations and rotations which are sufficient to describe the movement of solid objects. We propose to register CT volumetric datasets to corresponding 2D X-ray fluoroscopic images. As a single 2D image, in practice, does not convey sufficient information about the spatial location of the imaged object, we require two projection images to achieve our task. We assume that the two imaging views are related by a known transformation. The main challenges of the problem lie in identifying a similarity measure that can quantify the quality of the alignment between the images and in defining a procedure to modify and refine current estimates of the transformation parameters in a way that the similarity score is optimized.

Motivation
Although 2D images lack significant information that is present in 3D modalities, they might be conveniently and efficiently used to record details about the most current state of the imaged object. The most amount of information can be gained about the changes recorded by the 2D modalities and the detailed 3D model if we fuse the information provided by both of them. In order to achieve the proper spatial alignment of the different components, it is necessary to determine their relative position and orientation.

Approach
We propose an intensity-based registration algorithm using an information theoretic objective function, mutual information (MI), to establish the proper alignment of the input datasets. For optimization purposes, we compare the performance of the non-gradient Powell method and two slightly different versions of a stochastic gradient ascent strategy: one using a sparsely sampled histogramming approach and the other Parzen windowing to carry out probability density approximation. In order to compare the multi-dimensional images, with the current estimate of the transformation parameters we create a simulated version of the 2D acquisitions. A pair of such Digitally Reconstructed Radiographs (DRRs) is compared to the observed X-ray images. Our main contribution lies in adopting a stochastic approximation scheme successfully applied in 3D-3D registration problems to the 2D-3D scenario, which obviates the need for the generation of full DRRs at each iteration of pose optimization. This facilitates a considerable savings in computation expense. We also introduce a new probability density estimator for image intensities via sparse histogramming, derive gradient estimates for the density measures required by the maximization procedure and introduce the framework for a multiresolution strategy to the problem.

The following link leads to animated gif files demonstrating the incremental image aligning procedure. (animation)




FIGURES: Registration results of an experiment on X-ray and CT acquisitions of a skull dataset. Contours of the DRR images created by the output of the registration algorithm are overlaid on the observed fluoro images. (a)-(b) DRR contours created using the initial transformation estimate; (c)-(d) DRR contours created using the transformation estimate resulting from the registration process.

Publications

Lilla Zöllei: 2D-3D Rigid-Body Registration of X-Ray Flouroscopy and CT Images.
Masters Thesis, MIT AI Lab, August 2001. [PDF] [Postscript]

Lilla Zöllei, Eric Grimson, Alexander Norbash, William Wells:
2D-3D Rigid Registration of X-Ray Fluoroscopy and CT Images Using Mutual Information
and Sparsely Sampled Histogram Estimators.
IEEE CVPR, 2001, to appear. [PDF] [Postscript]


Researchers

William Wells                 sw@ai.mit.edu

Alexander Norbash       anorbash@partners.org

Lilla Zöllei                       lzollei@ai.mit.edu


Back to the Medical Vision Group page.

Back to the MIT AI Lab page.


Last updated September 21, 2001.
Lilla Zöllei