Biomedical Engineering Reference
In-Depth Information
they rely on the need for a priori knowledge of the regions that are affected by
a disease, so that respective ROIs can be defined, and therefore they might fail
to discover new findings. Although a good hypothesis might be available in the
beginning of a morphometric study, one would typically want to discover new
knowledge that, by definition, is not part of the hypothesis. As an example selected
from the neuroimaging of dementia literature, although the role of hippocampal
and entorhinal cortical atrophy in early prediction of Alzheimer's Disease (AD) is
widely accepted, relatively little is known about the potential involvement of other
brain regions, which could help construct more sensitive methods for detection of
and differentiation among different types dementia. The complete investigation of
the role of all brain structures in a disease and its diagnosis would be prohibitively
labor intensive for an adequately large sample size, if manual methods are em-
ployed. Moreover, inter- and intra-rater reliability issues would become crucial
limiting factors, particularly in longitudinal studies in which it is extremely dif-
ficult to maintain intra- and inter-rater reliability over time. Second, the spatial
specificity of ROI-based methods is limited by the size of the ROIs being mea-
sured, which is typically rather coarse. A region that might be affected by disease
may be only part of a pre-defined ROI, or it might span two or more ROIs, which
inevitably washes out the results and reduces the statistical power of the measure-
ment method. Alternative methods, such as stereology, are also limited in a similar
way. Although in principle one could define the size of the ROIs measured to be
as small as desired, in order to increase spatial specificity, this would decrease
rater reliability for measurement methods that are based on human raters. Finally,
manual ROI tracing is severely limited in many modern studies, for which it is not
unusual to include over a thousand scans per study.
In order to address the limitations of ROI-based approaches, image analysis
methods based on shape analysis have been studied in the literature during the
past 15 years [23-38]. One very promising approach for morphometric analysis is
based on
shape transformations
, and the associated methods are often called un-
biased, or hypothesis-free methods. A shape transformation is a spatial map that
adapts an individual's brain anatomy to that of another. The resulting transforma-
tion measures the differences between the two anatomies with very high spatial
specificity, and ultimately the specificity allowed by the image voxel size. More
generally, a template of anatomy is first selected, which serves as a measurement
unit. The shape transformation that maps other brains to the template is determined
via some sort of image analysis algorithm, and it is used as a means of quantifying
the individual anatomies. Inter-individual comparisons are then performed by ap-
plying standard statistical methods to the respective shape transformations. Voxels
that display significant group differences or longitudinal changes are grouped into
regions. Therefore, there is no need to define ROIs in advance. Instead, the ROIs
are determined retrospectively via the voxel-wise statistical analysis of the shape
transformations. The concept of this approach is shown in Figure 1a, which is
based on some of the second author's earlier work on the corpus callosum [25].
