Image Registration Methods in MRI - Advance Image Processing in Magnetic Resonance Imaging - page 206

these artifacts [40] are also likely to remove some of the true fMRI signal.
Moreover, in fMRI registration some residual artifacts also can remain after an
“exact” image realignment [41], due to the latency of excited tissues when they
return to their original state. Both of these problems are yet unsolved.

7.7.2 M APPING OF F MRI ON A NATOMICAL I MAGES

AND B RAIN A TLAS

As previously described, the spatial normalization (i.e., the registration of the
subject brain image into a standard coordinate space) is a common task in fMRI
imaging. The use of this procedure can be also extended to other applications,
such as those in which data coming from different subjects needs to be compared.
In the case of spatial normalization, the registration problem implies the use
of an elastic transformation, whereas the registration metric can be based on a
feature extraction (also known as label based) or a voxel-based (nonlabel based)
approach. In the feature extraction approach, some anatomical features are
extracted from the atlas. These have to be extracted from the subject image, and
the registration is performed finding the best transformation that superimposes
the two feature sets. Landmark points can be manually defined, but the process
is error prone and time consuming, and requires the presence of an expert operator.
Surfaces can be better identified in brain and can be automatically or semiauto-
matically extracted. Voxel-based methods use the similarity measures previously
described: the sum of squared differences, correlation, and MI.
To reduce the computational complexity of the elastic registration algorithms,
a multiresolution approach is often adopted in spatial normalization [42]. In this
approach, only a few parameters are determined at a certain resolution. As
example, the entire data volume is used to describe global frequency deformation.
The volume is then split in some subvolumes and the local frequency deformations
are calculated for each subvolume. The process is iterated until the needed
precision is achieved. Another approach involves the reduction of the registration
parameters to a small number (e.g., 9 to 12 parameters). This loss of precision
in registration accuracy allows the registration procedure to be performed in a
reasonable time. Moreover, in f MRI studies, different subjects can present dif-
ferent patterns of the brain anatomy, so high-resolution spatial normalization
appears to be unnecessary in many applications. The main problem in elastic
registration is the choice of constraints or priors that must drive the registration
process. Priors are usually incorporated by means of some Bayesian approach,
using estimators such as Maximum A Posteriori (MAP) or minimum variance
estimate (MVE).

7.7.3

F MRI R EGISTRATION E XPERIMENT

To show the importance of the registration operation in f MRI, the previously described
procedure is applied to a data set of f MRI images acquired during a finger-thumb
tapping experiment in which the subject is asked to touch the index finger of the right