Biomedical Engineering Reference
In-Depth Information
to the change in the rotation angle as revealed by the sharp peak in Fig. 4.7.
This is expected since a slight change in rotation angle can amount to a very
large offset in pixel positions if the distance from the rotation center is large.
It is very important to have an accurate estimation of the rotation angle in
registration.
4.3.3
Success Rate, Speed, and Accuracy
Our implementation of mutual information maximization for retinal image reg-
istration works relatively well. Figure 4.9 shows the registration results of
Figs. 4.2a and 4.2b (temporal registration). The registration images can be dis-
played side by side (not shown here). Figure 4.9a shows the two registered
images in a checkerboard format, where the size of each checkerboard can be
adjusted. Figure 4.9b shows these two images in a moving curtain format, where
the vertical line can be moved left or right to check the continuity of the image
features (vessels) across the dividing lines. The lines can also be horizontal. To
the left of the line is the reference image. The part of the reference image to
the right of the line is clipped out by the line and not displayed. To the right of
the line is the registered floating image. The left side of the matched floating
image is also clipped out by the line and is not displayed. Figure 4.9c shows
these two images in an overlay format, where the alpha is 0.5 so one can see
one image through the other image. The maximum, minimum, average, absolute
difference, color composition, and other fusion methods are also implemented
in our software. The color composition selectively extracts the color channels
and assigns them to the composite image which is very useful when one inspects
the registration and presents the fusion results of color images (for example,
red-free and angiograph).
In this section we compare our results against Ritter
et al
. [5] in terms
of success rate, registration speed, and registration accuracy. They used the
simulated annealing as the optimization routine. Their program was written in
C and their results were obtained on a Pentium Pro 200 running Linux. Our
program was written in Java (JDK 1.4.0) and the results were obtained on a
Pentium IV running Windows XP. We also ran the program on a Pentium 233 run-
ning Windows 95. Their success rate is 100%. On average it takes 1240 seconds.
In their implementation they used the nearest neighbor interpolation first and
then bilinear interpolation in the last layer of iteration. As one would expect, if
