Image registration: topology preservation

One challenge in developing image registration algorithms constraining the estimated deformation to preserve topology. Here are two slices of 3D X-ray CT scans of the thorax of a patient at exhale (left) and inhale (right). Notice how the diaphgram moves down during inhalation.

If we apply conventional unregularized 3D image registration using B-splines to warp the inhale images onto the exhale images then we would get the following results. The grid shown on the right illustrates that the estimated deformation is unrealistically irregular, and not locally invertible. This is physically implausible.

Graduate student Se Young Chun developed an improved image registration algorithm (to be presented at ISBI 2008) that constrains the deformation to be locally invertible by including a special regularization function in the cost function that is optimized. His method is faster than the Jacobian penalty functions used previously and yields the following results. The estimated deformation is much more realistic for breathing motion.


CT of thorax at exhale and inhale, and estimated 3D deformations with and without invertibility constraint.


Back to Fessler home page