Biomedical Engineering Reference
In-Depth Information
in quality control, it may appear that automatic registration followed by
visual inspection is no better than direct, interactive registration. A useful
advantage, however, in this division of duties is that the human visual system
appears to be good at recognizing a very bad registration, but not so good at
finding a very good one.
34
Many registration systems rely on optimization pro-
cedures that will perform properly for the large majority of patients but will
occasionally fall into a nonglobal optimum that produces a bad result. The
distribution of this latter mode is typically far wider than that of the properly
working system, thus greatly increasing the overall variance in TRE. Visual
inspection can detect very bad registrations, effectively truncating the distri-
bution at some detection level. By catching these large outliers, visual inspec-
tion can return a system nearly to its correctly working state. As a result of
efforts to quantify the accuracy of visual inspection as a means of failure detec-
tion in intermodality registration, it appears that when suitable interactive
image viewing software is available, the human visual system can be relied on
to detect a TRE greater than 4 mm for MR-to-PET
27,35
and 2 mm for MR-to-
34
CT.
Thus, for automatic systems that tend to remain below these levels
when working properly, the benefit of visual inspection should be well worth
the effort.
6.3.2
Registration Circuits
A simple consistency measure has recently been proposed when at least three
images of the same modality, A, B, and C, are to be registered.
36-38
By indepen-
dently registering A to B, B to C, and C to A in what might be called a regis-
tration “circuit,” it is possible to follow a target point from A to B to C, and
back to A. If the TREs for the three registration processes are uncorrelated,
then the RMS (TRE) for a single registration will be equal to RMS (TRE)/
for the circuit. For an
n
-image circuit the ratio is . Here, in analogy to Equa-
tion 6.3, the relationship between the circuit error and the error for a single
registration provides a bridge between consistency (circuit TRE) and accu-
racy (single-registration TRE) when registration errors are uncorrelated. This
approach has been proposed for serial MR registration problems, where it may
be important to measure TRE values as small as 0.2 mm
38
or even 0.05 mm.
12
The circuit may be the only approach available for such applications in view
of the difficulty of producing a gold standard with this extreme accuracy. The
same approach can be used for intermodality registration if two images of
each modality are available. One needs only to arrange the images so that
modalities alternate around the circuit.
While this approach is reasonable when no gold standard is available, a
potentially serious weakness lies in its assumption that the errors between
registrations are uncorrelated. The assumption will inevitably be violated
because of the fact that successive registrations, A to B and B to C, say,
share a common image. Noise or imaging artifacts in B will tend to cause B
to be shifted in the same direction relative to A in the A-B registration as it is
3
n
