Biomedical Engineering Reference
In-Depth Information
fiducial systems have the advantage over the frame in that they do not rely
on maintaining a prescribed shape among the fiducials. Thus, errors caused
by subtle bending of the
-bars or head ring of a frame are not a problem for
fiducial systems. Fiducial systems are, however, like the frame, subject to
errors caused by relative motion of the marker and anatomy between image
acquisitions, or between image acquisition and physical localization. Such
problems may be severe for skin-attached markers. The use of large numbers
of skin markers enhances their accuracy only marginally because of corre-
lated FLEs as they ride together on moving skin. Bone-implanted fiducial
marker systems are far less prone to consistent motion and therefore gain the
advantage of the bootstrapping feature, explained above, in which accuracy
can be inferred without independent validation. Thus, point-based valida-
tion systems based on bone-implanted markers provide self validation as
well. This benefit derives from the known statistical relationships among
FLE, FRE, and TRE, as described above, which follow from the independence
of FLE among the markers. Thus, while FRE is untrustworthy as a direct mea-
sure of registration accuracy, it can be exploited as an indirect measure if the
theoretical chain is followed correctly from its experimental measurement via
the estimate of the distribution of FLE to the prediction of the distribution of TRE
(see Chapter 3). With this theory in hand, it is possible to provide statistical
error bounds on the estimate of accuracy provided by point-based validation.
The point-based standard is superior to the target-marker approach because
it provides a dense set of TRE estimates, one at every point in space, whereas
the target-feature approach provides an estimate at only the points occupied
by features. Thus, for rigid-body registration in
N
M
dimensions, if there are
M
or more features available, it makes sense to use the point-based registration
approach. This approach also typically interpolates feature localization
errors by incorporating a single-least-squares fit for all of them.
6.3.1.4
Other Standards
Because a gold standard is simply a system whose accuracy is known to be
high, any registration system with a known error may in principle be used as
a gold standard. Wong et al., for example, used a mutual information method
as a gold standard for validating accuracy of the human visual system in the
detection of PET-to-MR misregistration. The visual system has been used in
countless reports as a gold standard for evaluating automatic registration
methods. It can play an important role during the development stages
of any automatic registration system as a means to assess its potential and
guide its refinements. It is also clearly applicable whenever the goal of the
system is to produce automatic registrations that are on par with those that
could be done manually. The purpose of the automatic registration in that
case is to substitute inexpensive computer time for expensive expert time.
Visual inspection for validation can perhaps serve best by providing qual-
ity assurance in clinical use as a safeguard against harmful errors for a given
patient. Because visual inspection has the last word on accuracy when used
27
23,28-33
