Nonrigid Image Registration

Nonrigid registration is the process of finding a coordinate transformation between two or more images that is anatomically correct and not constrained to be rigid or affine. Determining such transformations, which frequently have several hundred thousand degrees of freedom, requires substantially more computational effort than finding a rigid transformation.


Magnetic resonance (MR) brain image. Left: before surgery. Right: during surgery. (Figure from [2]).

Our work on nonrigid registration aims to develop new registration methods that improve the performance of current algorithms and increase the accuracy of the computed transformations. We apply high performance computation hardware (shared-memory multiprocessor architectures) to execute highly parallelized computations. Using novel numerical regularization techniques we incorporate prior physical knowledge to improve the registration results and make them more accurate descriptors of the actual physical situation.


Subtraction of pre- and intra-operative brain image. Left: after affine registration. Right: after nonrigid registration. (Figure from [2]).

One common application of nonrigid registration algorithms is to compensate for motion that occured between two image acquisition from the same subject (e.g., patient). For example, brain shift during cranial surgery causes deformations between pre- and intra-operative images, which need to be determined and taken into account when using pre-operative data for navigation and targetting during surgery.

Another application is the registration of images from several different subjects. Such inter-individual registration can be used, for example, to register an image to a segmented atlas image and thus obtain a segmentation of the new image. We also use nonrigid registration to generate statistical models of shape differences between individual and between groups of individuals.

  1. T. Rohlfing, C. R. Maurer, Jr., D. A. Bluemke, and M. A. Jacobs, ``Volume-preserving nonrigid registration of MR breast images using free-form deformation with an incompressibility constraint,'' IEEE Transactions on Medical Imaging, vol. 22, pp. 730-741, June 2003.
  2. T. Rohlfing and C. R. Maurer, Jr., ``Nonrigid image registration in shared-memory multiprocessor environments with application to brains, breasts, and bees,'' IEEE Transactions on Information Technology in Biomedicine, vol. 7, no. 1, pp. 16-25, 2003.

URL: http://www.stanford.edu/~rohlfing/research/nir/index.html
Last updated October 19 2009 11:45:29.
Torsten Rohlfing, Ph.D., torsten@synapse.sri.com
SRI International, Neuroscience Program
333 Ravenswood Avenue, Menlo Park, CA 94025-3498, USA