Biomedical Engineering Reference
In-Depth Information
13.2.2.3
Accuracy
The accuracy of all 41 experiments with respect to the fiducial-mark-base
gold standard can be found on the web (see http://www.vuse.vanderbilt.edu/
image/registration/).
In addition, we compare the results of our approaches to those of sev-
eral other approaches published in the literature. The comparison is based
on a methodology proposed by West and Fitzpatrick [21], who let selected
researchers access a standard set of image volumes to be registered. They
also act as a repository for the ideal registration transformations (gold stan-
dard) acquired by a prospective method using fiducial markers. These mark-
ers are removed before the volumes are disclosed to the investigators, who
then face a retrospective blind registration task. After registration, they email
back a set of transformation parameters that are compared to the gold stan-
dard.
Tables 13.1 and 13.2 show the median and maximum accuracy reached by
the investigators taking part in that project [21]. All errors are in units of mm.
The method using binarization approach is labelled LO1, while the method using
nonlinear binning technique is labelled LO2. We observed that both techniques
give impressive results for CT-MR registration.
13.2.2.4
Optimization Steps
In the downhill simplex optimization method, the number of optimization steps
is measured by the number of times the cost function is called. The mean number
and the standard deviation of optimization steps for the three binning techniques,
for each patient data set, is compared for CT-MR registration in Table 13.3 and
Table 13.4.
The binarization approach needs the least number of optimization steps in the
first level (Table 13.3). The methods using linear binning and nonlinear binning
need 30% to 102% more steps than the binarization approach. Of the three binning
techniques, the binarization approach has the most stable performance for all
the patients. The reason is that binarization can give an extreme blurring of the
images and so eliminates local optima. The performances of linear and nonlinear
binning are pretty much the same, while nonlinear binning is better in four out
of seven patients.
