Biomedical Engineering Reference
In-Depth Information
Another complication associated with the deformable registration of brain
tumor images is the significant signal changes associated with edema in MR
images. Edema typically causes hypointensity changes in T1-weighted images,
which makes it difficult to discern cortical sulci in the affected brain regions. It is
therefore not possible to obtain an accurate deformable registration in these regions
based on image matching alone. The coordinates of brain structures masked by
edema, however, may be estimated from the known structures (outside the edema
region) through a statistical estimation approach [98].
Dealing with the small sample size problem for statistical learning in very
high-dimensional spaces is another challenging problem that needs to be addressed
in future work. This problem made it necessary to collect statistics over part, and
not the whole, of the atlas domain outside
M
A
. Rather than restricting the region
of the atlas space where statistics are collected, an alternative approach to dealing
with statistical learning problem in high-dimensional spaces from a small sample
size was presented in [103]. This approach collects the statistics via PCA over
subspaces constructed from a wavelet packet basis, and that correspond to different
levels of detail of the analyzed shape or deformation.
5.
CONCLUSION
Voxel-based morphometric analysis provides an unbiased way of examining
high-dimensionality image data, in that it does not rely on any a priori hypotheses
regarding the regions of interest to be examined, but it analyzes the entire image
dataset and identifies regions in which morphological measurements are of interest
(e.g., regions in which volumetric measurements differ between normal controls
and patients, or between two serial scans). There is a plethora of voxel-based
morphometric analysis methods [104], each of which has merits and limitations.
In the first part of this chapter, we described a technique based on high-
dimensional elastic warping of brain images, formulated in a mass-preserving
framework so that tissue volumes are properly preserved and measured in this
process. We also discussed advantages of high-dimensional pattern classification
and multivariate analysis over voxel-based (mass-univariate) methods, since the
former capture complex associations among imagemeasurements in different parts
of the brain. We also presented a few representative studies in which tissue density
statistical analysis was used as a means for volumetric analysis.
In the second part of this chapter, we presented an approach for deformable
registration of brain tumor images. Solving this deformable registration problem
makes available methods for brain image morphometry able to assist in quantify-
ing and understanding changes in brain structure due to the presence of a brain
tumor. In addition, deformable registration between brain tumor images and a
brain template will facilitate the construction of statistical brain tumor atlases that
will be instrumental in neurosurgery and radiotherapy planning, and in the study
of brain cancer.
