Biomedical Engineering Reference
In-Depth Information
It is highly desirable to have an automatic method for evaluating the quality
of a registration so that a poor one can be flagged before it is used clinically.
The correlation coefficient would be applicable whenever one uses MR images
obtained with identical pulse sequences. It compares favorably with the bony
landmark results. Registration consistency provides an additional means to eval-
uate registration accuracy that does not rely on operator interaction.
Other evaluation methods are applicable for clinical or research applica-
tions. RegViz provided visual inspection tools for quick evaluation of the quality
of registration and potential prostate displacement. Such methods can be used
to verify the quality of registration and possibly account for small displacements
in some applications. Boundary overlays provide a good means to evaluate or-
gan deformation as well as displacement. Point anatomical landmarks provide
a useful, independent test, but it is time consuming to identify them and MI
might be more accurate than the point landmarks. Centroids are obtained re-
liably because small segmentation errors are removed by integrating over the
entire prostate volume. Centroids provide a good means of quantifying prostate
displacements.
3.2.5.3
Algorithm with Combined Similarity Measures
Using both CC and MI at different resolutions was an important feature that
increased robustness. When only mutual information was used, registrations
at low resolution sometimes gave false solutions that mislead registration at
the next higher resolution. However, CC performed well and gave many fewer
local maximums at the lower resolutions (Figs. 3.1a and 3.1b). But MI gave
a more accurate solution at the full resolution due to the peaked MI surface
(Figs. 3.1c and 3.1d). Our registration algorithm combined advantages from the
two similarity measures.
There are probably several reasons why mutual information does not work
well at low resolution. First, the similarity curve is noisy with periodic oscilla-
tions from the so-called interpolation artifact [30] that is accentuated at reduced
resolutions [35]. This results in the many local maximums in Fig. 3.1a that can
trap the optimization. A similar result was reported for brain registration [19,
36]. Second, when images are of low resolution and there is only a small region
of overlap, the mutual information function can even contain incorrect global
maximums [35]. Such a result was found in Fig. 3.1a where the global maximum
