Biomedical Engineering Reference
In-Depth Information
position of two images is sufficiently close to the optimal alignment can they be
registered well. To increase the capture range, naturally the next step is to exploit
mutual information. Due to its super performance, mutual information becomes
the first choice for automatic image registration. The advancement on mutual
information registration has been reviewed recently by Pluim
et al
. [22].
4.1.1
Review of Retinal Image Registration
Retinal image registration is the main focus of this chapter. This registration gen-
erally involves large x translation, due to changes between sittings and smaller
y translation from changes in position of the chin cup. Rotation occurs due to
tilting of the head and through ocular torsion, and scaling is caused by changes
in the distance between the camera and the head, due to equipment changes
or differing head positions (see [5, 23]). This section reviews some registration
methods as applied to retinal images. By no means is this review complete. The
interested readers may refer to [3, 5, 23] and references therein for more related
work.
Peli
et al
. [24] reported on a correlation method that preprocesses the images
using an adaptive threshold procedure to select vessel points. The normalized
sum of differences is then calculated with these vessel points. Via an exhaustive
search, this method produces pixel-level registration for
x
- and
y
-translation
only. It is not robust toward large changes in image intensity and white noise. The
absolute value of difference of pixel intensities was also used as a comparison
measure for retinal image registration. The images can be processed twice, using
optic discs as features for coarse alignment and the blood vessels as features
for fine alignment [25].
If one image is a scaled, rotated, and translated version of another image,
then the Fourier transform of that image is a scaled and rotated version of the
Fourier transform of the other image. Thus, image registration can also be done
in Fourier transformed space. Cideciyan
et al
. [26] computed the scaling and
rotation differences of the Fourier transformed images by cross-correlation.
These results are then used to transform one image in spatial domain. The final
translation differences in spatial domain are then found via cross-correlation.
When the images are taken at different times, where the translation difference
and the image intensity difference may be large, this approach is problematic.
This approach is not applicable to multimodality image registration.
