Biomedical Engineering Reference
In-Depth Information
work well - even if, as in Figure 3.17, one colors the two images
differently (e.g., red and green, so that the overlap regions appear
yellow). In the special case of two images of the same modality, the
same orientation and the same scale, this registration technique is not
too difficult. But, as is almost always the case, if the images have
different scales, or are rotated relative to one another, or are from
different modalities (e.g., CT and MRI, or radiograph and DRR), this
approach is near-hopeless. The overlapped image is simply too con-
fusing to the eye - and there are too many interacting variables to
make accurate manual registration feasible.
In practice, there are three main approaches to rigid body image
registration in both 2D and 3D, as now described.
Point-to-point registration
This approach is based on the
ability to identify of a pair of points, one in each section, which mark
the locations of the same anatomic or fiducial feature. This is done
multiple times for multiple anatomic features. If at least three non-
collinear point pairs are identified, then the rigid body transformation
between the two studies, namely the translations and rotations needed
to bring them into alignment, can be calculated mathematically. That
is the technical way of saying that, given the location of a feature in
one image, one can compute the location of the same feature in
the other image. While three non-collinear point-pairs are mathe-
matically sufficient for reconstruction, the solution is much more
robust - that is, is much less sensitive to errors in feature localiza-
tion - if a larger number of features are identified and a least-squares
fit is made to all the point-pairs.
Surface-to-surface registration
There is a very profound problem
with point-to-point registration. Namely, the human body really
doesn't have anatomically distinct “points.” It is composed of
volumes of soft tissue or bone whose boundaries are delimited by
surfaces, not points. The idea that anatomic
points
can be identified is
intrinsically wrong and the attempt to do so is prone to error. A much
sounder approach is to match anatomic
surfaces
with one another.
One of the first surface-matching algorithms, developed in the early
days of automated image registration, was the so-called “hat and
head” model which matched the inner table of the skull in a pair
of imaging studies (Pelizzari
et al.
, 1989). In the case of inter-
registering a pair of 2D projection images, the surfaces of the various
volumes such as skin, bone, airways, and so forth appear as
curves
and the task is to match pairs of curves with one another. This may
