Biomedical Engineering Reference
In-Depth Information
of the performance of a registration system may also be considered in deter-
mining its success. Maintz and Viergever, for example, in a thorough anno-
tated survey of papers on registration (through 1996) list precision, accuracy,
robustness
stability, reliability, resource requirements, algorithm complexity,
assumption verification, and clinical use as the items to consider as part of the
assessment.
1
Our choice of the degree of alignment as the single measure of
success simplifies assessment of registration methods and comparisons
among them. A major benefit is that no patient outcomes are involved in this
measure of success. Measuring registration success on the basis of patient
outcomes necessarily convolves the quality of the alignment with the quality
of the operation of each component of the embedding system through the
chain from registration to diagnosis or treatment. In fact, such derived mea-
sures of success, in which improvement in registration accuracy is deduced
from improved output of some subsystem further along the chain, may be all
that is available when alignment cannot be measured directly. This derivative
approach has been applied, for example, in the registration of bilateral
mammograms by assessing the visual detection of abnormalities.
2
Such val-
idations, which are capable of certifying a registration system as acceptable
or of ranking systems, replace the problem of registration assessment with
a different assessment problem. They are, however, to be avoided, if possi-
ble, in favor of the direct measure. The resulting simplification of the assess-
ment report and sharper definition of registration quality are well worth the
effort.
Most of the work and much of the literature on the subject of registration
inevitably focuses on the quest for registration methods that produce a better
alignment for some combination of modalities. The success of the registra-
tion, which we are relating monotonically to the quality of the alignment, has
been estimated in published work by visual inspection, by comparison with
a gold standard, or by means of some self-consistency measure. Although the
great majority of studies of registration quality have been carried out for
rigid-body registration algorithms, the same concepts are also applicable for
nonrigid registration. The estimates of registration quality are employed in
two distinct notions of success and, concomitantly, of failure. The first is suc-
cess for a class of image pairs; the second is success for a given image pair.
The former measure is useful when determining whether a method is appli-
cable for a given clinical problem, while the second is useful as a safeguard
against harmful errors for a given patient. In either case it is necessary to
determine alignment error.
6.2.1
Alignment Errors
The measurement of registration success, whether for a class of images or for a
given pair of images, will be some statistical estimate of some geometrical mea-
sure of alignment error. Many such measurements have been used to measure
the quality of registration, but not all are of equal value. An understanding of
