Description of Registration Method for Ashburner
The coregistration is a three step procedure:
1) Simultaneously estimate 12 parameter affine transformations that register
the two images to template images of the same modalities and contrasts.
A Newton-like least squares method can be used since the images and templates
are of the same modalities. The affine registrations are constrained so
that only parameters reflecting rigid body displacements are allowed to
differ between the pair of affine transformations.
2) The method assumes that the brains are composed only of three tissue
types: gray matter, white matter and CSF. Gray and white matter partitions
are extracted from the images using a modified mixture model cluster analysis
method. The model assumes a spatially varying a priori probability of each
voxel being a particular tissue type. The prior probabilities are described
by images conforming to the space of the template images described in
step (1). Since the mappings between the images and templates is known,
then the a priori images can simply be ovelayed on to the images to be
partitioned. This partitioning is an iterative procedure, whereby parameters
describing the distribution of voxel intensities within each partition are
computed. Using these parameters and the a priori probability images, a
probability is determined for each voxel belonging to each of the
partitions. The loop continues by recomputing the parameters of the
distributions using the probabilities, and using the probabilities to
recompute the parameters.
3) The affine transformation parameters obtained from step (1) are used as
starting estimates for the final step, which involves registering the image
segments extracted using step (2). Rigid body transformations are estimated
that simultaneously match both the gray and the white matter partitions
from the pair of images, by minimizing the residual squared difference
between them.
Method 1 uses only the first step. Method 2 uses all three steps.
References
1. "Multimodal Image Coregistration and Partitioning - a Unified Framework.",
J. Ashburner & K.J. Friston, 1997, NeuroImage, Vol. 6, Issue 3,
pp. 209-217.
Return to previous page