Biomedical Engineering Reference
In-Depth Information
4.3.3.2
Registration of Stereo Images
Visual inspection of the registration results reveals that mutual information max-
imization registration succeeded 9 times out of 11 and it failed on ll5/lr5 of set 1
and ll7/lr7 of set 3. The success rate is 82%. The average registration time is 11.0
seconds with a standard deviation of 1.5, which amounts to a speedup of 8.39.
Again, the speedup value is only indicative.
For the stereo registration, the mean and standard deviation of misregistra-
tion parameters are (
−
0.46
±
1.69,
−
0.67
±
1.72,
−
0.02
±
0.24, 0.0043
±
0.0082).
The estimation of the rotation angle and the scaling factors are very accurate.
The RMSE numbers are (1.75, 1.85, 0.25, 0.0090). The translation RMSE num-
bers are larger than what Ritter
et al
. reported, but the rotation and scaling are
comparable to their results. As we mentioned earlier, it is hard to pair the points
in two stereo images.
4.3.3.3
Registration for Temporal and Stereo Images
If we combine all results discussed above, then the success rate is 86%. The
average time for registration is 11.7 seconds with a standard deviation of 2.6,
which is about 7.83 speedup against Ritter
et al
. [5].
The average misregistrations are (
−
0.32,
−
0.17,
−
0.01, 0.0021) for
x
and
y
translations, rotation angle, and scaling factor, respectively. The RMSE are (1.49,
1.39, 0.21, 0.0070).
4.4
Summary
Retinal image registration and fusion is an important tool in ophthalmology for
diagnosis, lesion progression assessment, and treatment monitoring. Registra-
tion by mutual information maximization is one of the most popular automatic
algorithms. We implemented such a software system for automatic retinal image
registration by mutual information maximization. The software was written in
Java and Java 2D for maximum portability. The robust software architecture
(MVC framework) was employed to ensure software maintainability and ex-
tensibility. The simplex downhill method was exploited to optimize the mutual
information function as a function of the registration parameters. To increase
the search range and to avoid being trapped by local optimal, multiresolution
