Biomedical Engineering Reference
In-Depth Information
As all images in an fMRI experiment are taken a few seconds apart, with the
same settings, on the same scanner, and with the same subject, it is an almost ideal
intrapatient, intramodality registration scenario. Consequently, rigid-body trans-
formations with intramodal similarity measures, as described in Chapter 3, are
almost universally used to model the change between one image and the next.
In addition to rigid-body motion, there are also some sources of nonrigid
motion usually present. For example, pulsatile motion of the soft brain tis-
sues occurs during the cardiac cycle. Even bulk motion of major chest organs
during respiration will change the magnetic field distribution throughout the
body (including the head) and will therefore affect the geometry of the scans,
inducing nonrigid motion. It is possible to reduce the extent of some of these
motions by using methods such as cardiac gating of the images (acquiring at
the same point in the cardiac cycle each time). However, the major compo-
nent of motion is due to rigid movement of the head in the scanner, and the
correction of this motion will be discussed in the following sections. For more
detail on physiological noise, see Hu et al. and Jezzard.
5,6
8.2.1
A Multiple Registration Problem
The basic problem of motion correction is to align each image in the series
to a consistent orientation (by registering each one to some fixed target image).
Therefore, motion correction is simply a series of registrations. However,
there are typically more than a hundred images to register, so speed is
quite important. For instance, if each registration took 30 minutes and
there were 200 images, the total motion correction time would be more
than 4 days.
The standard approach taken in motion correction is to use an intramodal-
ity voxel similarity measure, such as sum of squares of differences or mean-
absolute-difference, together with some optimization algorithm. The registration
can be performed quite rapidly, since the resolution of the images is relatively
low, the motion is usually small (so that simple local optimization will suf-
fice), and only 6 degrees-of-freedom (DOF) transformations are used (i.e.,
rigid-body). For instance, a 200-image series can typically be motion corrected
in less than 30 minutes—that is, less than 10 seconds per image.
The method of minimizing the cost function while varying the translation
and rotation parameters is usually a standard optimization technique (such as
gradient descent, Powell's method, or a simplex method—see Reference 13 for
more details) or a customized version of these. It is assumed that any activation
in the images does not affect the estimation of motion, and since the intensity
changes due to activation are relatively low, this is probably a safe assumption.
It is necessary to choose a “target” image for motion correction. Normally
this is simply an arbitrary single image from the original data. However, care
needs to be exercised when choosing the target, as the first image from the
sequence probably looks quite different from all subsequent images, due to
MR saturation effects. Therefore, if this is used as a target, the cost function
