Biomedical Engineering Reference
In-Depth Information
and subsampling optimization schemes were explicitly employed. Other imple-
mentation strategies were also tested, including nearest neighbor and bilinear
interpolation and histogram estimation. Our implementation has an 86% success
rate. Compared to the published results, our implementation is about 7.83 times
quicker with comparable accuracy (mean of misregistration parameters) and
precision (standard deviation of misregistration parameters).
Besides the improved registration performance, our system also provided
new representation tools to visualize the registration results. The registered im-
ages can be displayed side by side to allow direct comparison. Moreover, the
registered images can be displayed in a moving curtain fashion and in a checker-
board format, where parts of two images are displayed together. Furthermore,
two images can be overlaid to allow one to see one image through the other.
These tools were integrated seamlessly to allow the user to check and interpret
the registration findings.
We attribute our fast registration speed to the multiresolution and subsam-
pling scheme. To our knowledge this is the first time anyone has taken advantage
of the benefits of both in a single implementation. Depending on the image size,
the coarse image can be very coarse, thus the registration can be very fast.
While Ritter
et al
. reported a 100% success rate, we only achieved an 86% rate.
We found that the resolutions and subsampling frequencies in our multiresolu-
tion and subsampling scheme can be adjusted so that the failed registrations
can be registered successfully. Our implementation provides a facility to allow
the user to change them at runtime. Considering that the registration can be
done in less than 15 seconds, it is practical and acceptable to register the failed
image pairs with a different set of parameters. Another approach to achieve a
higher success rate is to bring the images close to the optimal alignment before
starting the automatic alignment. This prealignment proves to be useful in 3D
registration and our preliminary results indicate that it is helpful in 2D retinal
registration too.
Acknowledgement
We are grateful to Dr. Nicola Ritter for providing us the retinal images used
in this study. Thanks are also due to Dan Kovalik for his proofreading of the
manuscript.
