Biomedical Engineering Reference
In-Depth Information
6.3.1.1
Computer Simulations
Computer simulations are image pairs generated by computer. While it is
possible to model the anatomy and imaging process to produce both images
to be registered, computer simulations are typically produced by modifying
one acquired image to produce a second simulated image that has been
modified by a known geometrical transformation. The second image may be
of the same or different modality. Simulations have the primary advantage
that the transformation is known exactly, and the secondary advantage that
any transformation is readily available. They are especially apt for development
work, but they lack the realism arising from the sometimes subtle anatomical
changes that accompany respiration, the cardiac cycle, and involuntary or vol-
untary motion during image acquisition. Such motion can cause warping in
CT, in which slices are acquired serially, and blurring and other reconstruc-
tion artifacts in MRI, PET, or SPECT, in which information is gathered from
multiple slices in parallel. Motion can cause other artifacts such as changes
in magnetic susceptibility patterns in MR or scatter patterns in the other
modalities, each of which is difficult to predict and varies with patient posi-
tion and orientation. Validations based on simulations can, however, provide
an upper bound on success, which can be of great value when a registration
system is under development.
The simplest application of computer simulation is intramodality registra-
tion. If the second image is to be of the same modality as the first, the modi-
fication that produces it is purely one of geometrical transformation. Because
a transformed voxel position will rarely fall on an integral voxel position,
interpolation is necessary (see Chapter 3). If the resolution in the first image
is strongly nonisotropic, the simulated image interpolation artifacts may be
unacceptable in the generated image. For this reason, computer simulation
for registration validation is typically based on MR volume acquisitions,
whose resolution can be made virtually isotropic.
11-13
For intermodality reg-
istration it is necessary to identify tissue types and generate new gray level
values (in addition to the new values resulting from interpolation) according
to the physical processes of the second imaging modality. Collins et al.
13
have
produced one high-resolution image of the head segmented into four tissue
types and background expressly for this purpose.*
For intrapatient registration, it is most common to determine rigid-body
transformations which, as stated above, are appropriate in most cases for the
brain. In other intrapatient applications, nonrigid transformations may be
more appropriate than model changes in soft tissue due to natural motion,
resection, growth, atrophy, or other physical changes (see Chapters 13 and 15).
For interpatient registration the transformation is either affine or a nonrigid
transformation with additional degrees of freedom, and must model human
variation. Such transformations are discussed in Chapters 13 and 14. Simu-
lated images for validation in the nonrigid regime must be approached with
* This image and simulated images generated from it are available at http://www.bic.mni.
mcgill.ca/brainweb/.
