Biomedical Engineering Reference
In-Depth Information
some part of a patient's anatomy. We have focused on three-dimensional
views and adopted a generalized notion of a “view” which includes not only
images acquired with imaging systems (for example, CT, MR, PET, or SPECT),
but also the physical view of the anatomy (for example, the view during sur-
gery). This notion of views accommodates both intramodality and intermo-
dality registration, and intrapatient and interpatient registration as well. We
define several measures of error including
target registration error,
or TRE,
which is the disparity in the positions of two corresponding points after reg-
istration. Regardless of the views, our definition of registration leads to a rec-
ommendation of TRE as the quantity of choice to be reported in the validation
process.
TRE can be expected to vary with the registration situation, which comprises
the imaging modalities, anatomy, and pathology. Also, for a given registra-
tion situation TRE can be expected to vary with position within a view.
Experimental validation of a registration system should thus be extended to
a clinical situation only to the extent that the clinical situation matches the
experimental one. The degree of the required match will vary with the regis-
tration system, but the same modality pair should always be used.
The most commonly accepted strategy for validation is to compare the sys-
tem to be validated against a
gold standard
, which we define as any system
whose accuracy is known to be high. Gold standards may be based on com-
puter simulations (typically by acquiring one image and generating a second
with a known geometrical transformation,
phantom images), cadavers, or
patient images. Computer simulations provide arbitrarily accurate geometrical
transformations but, like phantoms, are less realistic than cadaver or patient
images. Simulations should also be approached with great care in nonrigid
validations because of the bias of such validations in favor of registration
methods that employ similar nonrigid transformations, whether or not they
are physically meaningful. Validations based on pairs of acquired patient
images represent the most desirable class of standards because of the inclu-
sion of all the physical effects of the patient on image acquisition, but suffer
from the difficulty of establishing the true transformation between acquired
(as opposed to simulated) images.
The simplest method for establishing the transformation between acquired
images, effective both for phantoms and patients, is based on the
target fea-
ture,
which is any object that can be localized independently in each view. The
root-mean-square (RMS) disparity in the two localizations of the target fea-
ture after registration provides an upper bound on the RMS of TRE at the
location of the feature. A more desirable method for rigid-body registration
is based on a registration system that employs several fiducial features as reg-
istration cues. The major advantage of this type of system as a validation
standard is that its accuracy can be determined without reference to other
standards. This is accomplished by exploiting theoretically established statis-
tical relationships among fiducial localization error
(FLE), fiducial registration
error
(FRE), and TRE to translate self-consistency into accuracy. FRE plays an
important role in this translation, but is a poor measure of registration error.
