Biomedical Engineering Reference
In-Depth Information
that the alignment is perfect, but it does mean that the alignment error is
below some threshold. While registration accuracy cannot be determined
directly for a given clinical case (otherwise, there must be a superior registra-
tion method available), there may be some measurement that can be made for
a given case to indicate whether the accuracy is likely to be acceptable. The
relationship between that measurement and a clinically meaningful measure
of accuracy, such as TRE, will be statistical and must be verified experimentally.
The function that is optimized during the registration process can serve this
purpose, but its statistics must be carefully established to give it predictive
power. The primary example is FRE for a point-based rigid-body system.
Experimental measurements of FRE can be used to determine the distribution
of FRE values for a properly working point-based system.
6
During a clinical
registration, FRE can be reported and compared with that distribution. If the
value is very large, at or above the 95% level observed experimentally, for
example, the system may not be working properly.
6.3
Methods for Estimating Error
The most straightforward method for estimating registration error is to com-
pare a given registration transformation with a “ground truth” transforma-
tion, one whose accuracy is high. Ground truth may be obtained from some
gold standard
registration system, which we define as any registration system
whose accuracy is known to be high. We devote most of this section to these
gold standard methods.
6.3.1
Gold Standards
Gold standards may be based on computer simulations, phantoms or cadavers,
or patients, where the latter category is understood to include normals as well.
Phantoms may range from rectangular blocks of plastic to realistic anthropo-
morphic models
10
and are second only to computer simulations in providing
known transformations. They are inferior, however, to patient images and
computer simulations in image fidelity. For this reason they cannot be con-
sidered seriously in registration validation except for development work or
as applied to the stereotactic frame, whose accuracy can be expected to be
largely unaffected by differences between the phantom and a patient. For
that reason we will discuss only this single application of phantom validation
(see Section 6.3.1.2). Similarly, because of physiological changes at death,
cadaver studies have limited use but are also appropriate for validating ster-
eotactic frames and, to a certain extent, any systems based on bone-
implanted marker systems. Gold standards based on patients are the most
difficult to establish. They may be based on target features, fiducial marker
systems, visual inspection, or other methods.
