Method for registering deformable images using random Markov fields

1. A method for registering a source image with a target image using multiple scales, wherein the images are deformable, the method comprising:

measuring dissimilarity between the source image and the target image;

using the measured dissimilarity for each scale of the multiple scales to minimize a discrete energy function, wherein the discrete energy function for each scale contains a data term and a smoothness term, the data term represents deformation vectors as discrete labels, and the smoothness term imposes prior probabilities on the deformation to preserve discontinuities, and wherein prior to warping the source image, each scale includes:

applying multi-scale Markov random field registration to determine a labeling which corresponds to a deformation vector field; and

accumulating the deformation vector field via deformation vector addition to an accumulated deformation vector field determined at a previous scale, wherein the deformation vectors are added on the basis that the deformation vector field at a final scale corresponds to the deformation vector field at a desired resolution, such that the final deformation vectors are aggregated at each scale prior to warping;

warping the source image, according to the accumulated deformation vector field at the desired resolution, to obtain the target image, wherein the steps are performed in a processor.

2. The method of claim 1 , wherein the dissimilarity is measured using a sum of squared differences.

3. The method of claim 1 , wherein the dissimilarity is measured using normalized cross correlation.

4. The method of claim 1 , wherein the dissimilarity is measured using mutual information.

5. The method of claim 1 , wherein the multiple scales are arranged in a coarse to fine order, and wherein the coarse scale uses integer pixel deformations, and finer scales use subpixel deformations.

6. The method of claim 1 , wherein the energy function includes the data term, and the smoothness term, wherein the data term is a measure of the dissimilarity, and the smoothness term is a regularizer.

7. The method of claim 6 , wherein the regularization term is imposed over accumulating the deformation vector field, such that local discontinuities in the deformation vectors are preserved.

8. The method of claim 7 , wherein the regularization term is spatially varying and determined locally from the input data.

9. The method of claim 1 , wherein the discrete energy function is minimized using a quadratic pseudo-Boolean optimization (QPBO) procedure.

10. The method of claim 1 , wherein the dissimilarity is obtained from a machine learning computation approach.

11. A system for registering a target image with a source image using multiple scales, wherein the images are deformable, the system comprising:

at least one processor, coupled to at least one memory, the at least one processor to perform the steps of:

measuring dissimilarities between the target image and the source image;

using the measured dissimilarity for each scale of the multiple scales to minimize a discrete energy function, wherein the discrete energy function contains a data term and a smoothness term, the data term represents deformation vectors as discrete labels, and the smoothness term is a regularizer that imposes prior probabilities on the deformation to preserve discontinuities, wherein prior to warping the source image, each scale includes:

applying discrete optimization to determine a labeling which corresponds to a deformation vector field, wherein discontinuities in the deformation vector field are preserved; and

accumulating the deformation vector field via deformation vector addition to an accumulated deformation vector field determined at a previous scale, wherein the deformation vectors are added on the basis that the deformation vector field at a final scale corresponds to the deformation vector field at a desired resolution, such that the final deformation vectors are aggregated at each scale prior to warping.

12. The method of claim 11 , wherein the discrete optimization includes using a quadratic pseudo-Boolean optimization (QPBO) in determining the labeling, and during use of QPBO, if a discrete label cannot be assigned for each pixel in the target image, an unknown discrete label is assigned, then an additional step is applied to assign an identified label that has corresponding energy that is lowest.