Nonparametric and nonrigid registration method applied to myocardial-gated SPECT

We have developed a nonparametric and nonrigid registration method. It requires a segmentation contour of the organ of interest in both the image to register and the image template. We use the fact that the contour template can be deduced from the contour in the image to register by a series of locally normal elementary deformations. We extend these contour deformations to the whole image domain using level sets (sets of embedded parallel contours, the organ contour being the contour of reference). This series of elementary deformations provides a way to perform an iterative (totally unrestricted, unlike parametric methods) registration by indicating the successive motions from a point in the image to register to its final position in the registered image. The use of the organ contour has two consequences: 1) Our method is consistent with the image features and 2) our method is based on image geometry (as opposed to intensity) and, therefore, it is less sensitive to noise and more robust in homogeneous regions than nonparametric methods based on optical flow. We have applied our registration method to perform a motion compensation of gated single photon emission computed tomography (SPECT) sequences, i.e., to nonrigidly register N - 1 frames with respect to the Nth one. The N registered frames were then added together to improve the signal-to-noise characteristics while minimizing motion-induced blurring. Results on three-dimensional (3-D) mathematical cardiac-torso phantom (MCAT) simulated sequences are presented.