Biomedical Engineering Reference
In-Depth Information
Hart
et al.
[27] exploited a blood vessel filter to detect the ends of blood
vessel segments and then used them as control points to register retinal images.
A specific scheme was designed to eliminate erroneous point pairs until either
there are four control points left or the mean square error given by the least
square fitting drops below five pixels. This method has some problems as pointed
out by Ritter
et al
. [5].
Ritter
et al
. [5] recently applied the mutual information maximization to
retinal image registration. To find a global optimum, simulated annealing is used
in the multiresolution optimization. Although this method can successfully find
the global optimum registration with translation, rotation, and scaling, it is time
consuming. To assess the accuracy of registration, Ritter
et al
. compared the
registration results against the solutions obtained by an exhaustive search. This
comparison has an intrinsic drawback. What they studied is how the simulated
annealing behaves which is an implementation artifact, not how the mutual
information maximization behaves as a registration criterion.
Matsopoulos
et al
. [28] used matched filters (see [29]) to segment the vessel
trees and registered the segmented trees automatically. To ensure that a global
optimal registration is found, simulated annealing and genetic algorithms were
employed. They also studied the suitability and efficiency of different image
transformation models. The criterion used in the optimization is a correlation
function defined on the segmented, binary images.
Zana
et al
. [3] reported on a multimodal retinal registration scheme based
on vessels detection and Hough transform. The vascular tree is segmented first,
and then the bifurcation points are detected. Those tree and points are features
used to register the images. Although their algorithm is attractive, it involves a
fair amount of user interaction in the preprocessing and in the final registration
selection (the solution given by Bayesian selection is not necessarily the best).
Laliberte
et al
. developed a similar technique that also used the blood vessel
bifurcation points, but did not need the assumption of a Gaussian shape vessel
intensity profile which is inappropriate for low resolution optical images (see
[23]). In spite of about 10 adjustable parameters in the algorithm, it seems that
the success rate for this latter method is low (36 out of 61 pairs). Can
et al
. [30]
developed a hierarchical scheme to match the feature points in two images, using
a progressively complex transformation model and a reduced set of matching
points. This algorithm is attractive when one builds the retinal map since warping
is generally required.
