Biomedical Engineering Reference
In-Depth Information
Table 4.1:
Comparison of Ritten
et al
. and our implementations of
mutual information maximization for retinal image registration
Ritter
et al
.
Our implementations
Operating system
Linux
Windows XP
CPU
Pentium Pro 200
Pentium IV, 2.4 GHz
Language
C
Java (JDK 1.4.0)
Interpolation
Nearest neighbor
+
bilinear
Bilinear only
Optimization
Simulated annealing
Simplex
Multiresolution
Yes
Yes
Subsampling
No
Yes
bilinear interpolation were used in all iterations, it would take longer. Table 4.1
summarizes the differences.
Ritten
et al
. reported the root mean square error (RMSE) of the registration
parameters. For the stereo registration, they are 0.59, 0.52, 0.31, 0.0058 for
x
and
y
translations, rotation angle, and the scaling factor, respectively; for the
temporal registration, they are 0.29, 0.29, 0.15, 0.0038, respectively.
For this kind of registration the ground truth is unknown. Therefore, it
is hard to evaluate its accuracy. Ritter
et al
. compared their results against
the results obtained from the exhaustive search in a limited area [5]. What
they assessed is not the mutual information maximization as a registration
method, but the simulated annealing as a global optimization scheme. It is
well known that the simulated annealing has a large chance to find the global
optimal solution. Rather than comparing the results against those of an ex-
haustive search, we compare our results against the point matching results.
In CT/MR/PET/SPECT image registration, point matching is regarded as the
gold standard [15], as long as one can identify the corresponding points
correctly.
It is possible to identify the corresponding points in retinal images if there
are enough vessel trees. Our images are about 5% of the total surface of the retina
centered on the optic nerve head. Some retinas have severe degradation. Some
images are blurred. The vessels are usually very thick. Thus the corresponding
points are not always easy to identify.
The point matching results of temporal image pairs and stereo image pairs
were obtained as the average of 4 or 5 independent registrations performed by
