Biomedical Engineering Reference
In-Depth Information
two operators. As we mentioned earlier, it is difficult to pair the correspond-
ing points in some image pairs. Thus we have less confidence in their results.
For example, in temporal registration, it is difficult to register the images in
Set 1. In general, the stereo images are difficult to register by point matching.
Nevertheless, the average registrations seem reasonable.
4.3.3.1
Registration of Temporal Images
Visual inspection of the registration results reveals that mutual information max-
imization succeeded 9 times out of 10 and it failed on lr6/lr7 of set 3. Thus the
success rate is 90%. The average registration time was 12.5 seconds with a stan-
dard deviation of 3.3.
We can compute the speedup of our implementation against Ritter
et al
. [5].
Considering the CPU difference and ignoring all other differences, we define the
speedup as
RittersTime
·
TheirCPUClockRate
OurTime
·
OurCPUClockRate
−
1
speedup
=
Substituting all numbers in, our speedup is 7.27. It is worth pointing out that
Pentium Pro and Pentium IV processors have different architectures, thus, this
comparison only has indicative meaning.
Define the misregistration as mutual information registration—“true” reg-
istration. The mean and standard deviation of these temporal misregistration
parameters are (
−
0.18
±
1.17, 0.33
±
0.56, 0.01
±
0.15, 0.0000
±
0.0040). Overall,
the estimation of the rotation angle and the scaling factors are very accurate. The
large errors in the translation offsets are primarily due to those errors related
to the first image set.
We also compute the RMSE and they are (1.17, 0.65, 0.15, 0.0040). Those
RMSE numbers shall be very much close to the standard deviation reported
above for a large dataset. The translation RMSE numbers are larger than what
Ritter
et al
. reported, but the rotation and scaling are better than their re-
sults. As we mentioned earlier, it is difficult to register image set 1. Excluding
set 1, the RMSE numbers are (0.34, 0.40, 0.06, 0.0020). The RMSE numbers of
the translations are comparable to those of Ritter
et al
., while the rotation and
the scaling factor are much better than theirs.
