Image Registration Methods in MRI - Advance Image Processing in Magnetic Resonance Imaging - page 205

Moreover, the spatial resolution of the two sets differs as well. A local elastic
transformation can also be applied to correct the geometrical distortions intro-
duced by fMRI.
A third important application of image registration in fMRI studies is the
warping of images from a number of subjects into a common standard space in
order to compare the data coming from different experiments on different subjects.
The most known standard data space is the Talairach atlas [34]. The warping of
the images into an atlas requires some elastic transformation.

7.7.1

F MRI I MAGES R EGISTRATION

Most current algorithms for motion correction in fMRI consider the head as a rigid
object, so that six parameters are needed to define the head as a rigid-body
transformation. The involved data sets are often anisotropic, and therefore an
appropriate transformation should be defined as previously described. Because
the fMRI signal is low, differences between images are small, and the sum of
squared differences between the image to be registered and the reference image
can be effectively used as a registration metric. This approach is used in SPM
and AFNI, the most well-known realignment packages for fMRI. However, the
use of more sophisticated metrics such as MI can avoid some artifacts produced
by the standard approach. Freire et al. [35] compared seven different motion
estimation procedures based on different choices of registration metric: two
different implementations of the difference of squares measure (SPM [36] and
AIR [37]); the Geman-McClure (GM) robust estimator [38], which takes into
account the existence of potential outliers; the ratio image uniformity function;
two symmetrical implementations of the correlation ratio (CR) [39], and MI.
Results suggested that GM, CR, and MI are robust in respect to the difference
of squares measure. Robust metrics such as GM or MI may be the best choices,
although in many applications the computational cost of these advanced metrics
may be too high in respect to the more simple algorithms required.
To improve registration quality, the registration operation is often followed
by a second registration in which all images are registered in respect to the mean
of all images realigned in the previous step. A variance image can be computed
together with the mean difference in order to provide a better weighting for
registration. In particular, the image voxels are weighted proportionally to the
inverse of the variance, so that voxels with a lower variance will have more weight
in the computation of the registration metric. As described in Section 7.6, more
sophisticated approaches can be applied to the registration of multiple data sets,
as happens in fMRI.
An important issue in fMRI experiments is the fact that subject motion is
often correlated with the experimental design. As an example, if the experimental
paradigm requires speech or hand motion, it is likely that head movement will
be highly correlated with the paradigm. In this case, it is extremely difficult to
separate the activation signals from correlated motion artifacts. In fact, methods
that use some measure of correlation between signal and paradigm to remove