Biomedical Engineering Reference
In-Depth Information
of the patients was also outfitted with stereotactic frames. All traces of the
markers and the frame
N
-bar were erased from the images before their post-
ing on the Internet, and the experimental situation was carefully documented
with regard to image acquisition. The anatomy included an approximately
consistent region of the heads of tumor patients. The frame, when present,
was used as a calibration device to correct for scaling errors in MR and, for
some patients, corrections were made for distortion arising from static field
inhomogeneity.
41
Retrospective systems were then applied to the images by
researchers at the institutions outside Vanderbilt and the resulting transfor-
mations were communicated, again via the Internet, back to Vanderbilt for
comparison with the gold standard transformations, which were sequestered
throughout the evaluations.
Erasure of the markers and sequestering of the standard transformations
blinds the user of the retrospective method to the correct transformations.
This feature is an important aspect of the study. It produces an assessment
more likely to predict clinical performance, because the system assessed can-
not be given inadvertent “hints” by the operator based on known answers.
The results of the study revealed that a median TRE below one millimeter is
possible for MR-to-CT rigid registration of head images, and a median of 3 mm
is possible for MR-to-PET. The study also showed that large errors sometimes
occur without warning, suggesting that some standard of quality assurance,
such as visual inspection, should be employed in clinical use.
These are important results, but they must be applied with care because of lim-
itations in the data set. The images came from a limited patient population
(patients with tumors that were about to be surgically resected) and include only
a limited variety of image acquisition protocols and scanners. Nevertheless, the
study has been of value in establishing with some certainty a benchmark for ret-
rospective registration accuracy, and the data set has been, and continues to be,
of value as a tool for developers. The evaluation service has been in continual use
since the original posting of the data set on the Internet in early 1996, both to sup-
port the work of developers
42
and to provide a means of independent validation
in published accounts of new methods.
43-46
* New patient data sets including
other regions of the anatomy, with new and better gold standards, would doubt-
less be welcomed by the image registration community.
6.6
Conclusions
Validation is an essential part of the registration process. It infers agreement
on the goal of registration, which we have taken to be the alignment of points
in one view of an object with corresponding points in a second view of an
object. The objects in these views are typically the same, usually a patient or
* Available at
http://cswww.vuse.vanderbilt.edu/~image/registration/
.
