Biomedical Engineering Reference
In-Depth Information
clinical meaning. Thus, the significance of angular motion is typically diffi-
cult to interpret clinically.
6.2.2
Statistics
As mentioned earlier, registration success may be determined for a class of
image pairs or a given image pair. The former measure is useful when
determining whether a registration system is appropriate for a given clini-
cal problem, while the latter is useful as a safeguard against harmful errors
during subsequent routine clinical use. Once a measure of alignment error
is chosen, it can be measured and reported, but because of the complex vari-
ation among images, it will be necessary to perform many registrations and
make many measurements before a clear picture of system accuracy can
emerge. The accuracy report is therefore necessarily statistical. Common,
meaningful statistics include the mean and standard deviation, the root-
mean-square, which is often used for Euclidean distance measures, the
median, which may be more meaningful than either of the former measures
in the face of outliers, and other order statistics, such as 90 or 95% thresh-
olds. Such statistics, gathered in experiments on a well-defined class of
images, can be used in predictions of clinical performance on future data
sets of the same class. Indeed, the primary purpose of gathering statistics
on experimentally measured image registration accuracy is the prediction
of clinical success or failure. For example, statistics gathered from experi-
mental measurements of TRE on a system for registering CT or MR with
specified imaging parameters (e.g., slice thicknesses, field-of-view, pulse
sequence, contrast administration, etc.) for a given anatomy (e.g., head,
liver, etc.) and given pathology (e.g., tumor, diffuse white matter lesions,
etc., or normals), can be used to predict the expected value of TRE, or pref-
erably its distribution, in a similar clinical situation. If the distribution is
available, then the probability of success can be stated as the probability
that TRE will fall below the acceptable threshold.
Except in rare cases, TRE will vary with position. Thus, the location, or
locations, at which TRE is measured is a critical part of the description of the
experimental situation as well. For example, if the clinical target is near the
optic nerve, then the value of TRE at the center of the cerebellum is not a reli-
able predictor of clinical success or failure. The mean TRE throughout the
brain is a valuable statistic but is still inferior to a point-by-point accounting,
even when it is supplemented by the variance. A maximum value provides
more protection against disaster, but such a conservative measure may pre-
clude the use of a system that would in fact benefit the patient for a particular
target point. It should perhaps be emphasized that experimental error statis-
tics are meaningful predictors of clinical error only when the clinical situation
is sufficiently similar to the experimental situation. We discuss this notion of
similarity further below.
A second, and equally important, purpose of gathering registration statis-
tics is detection of clinical failure. A successful registration need not mean
