Image Processing Reference
In-Depth Information
to have clustered activations. Spatial and temporal correlations affecting fMRI
signal measurements are typically not considered simultaneously in statistical
methods dedicated to detecting brain activation. In order to improve the detection
of activated areas, common approaches usually smooth the data spatially with a
Gaussian kernel in a preprocessing step. Spatial smoothing enables effective detec-
tion of a certain size of clustered activation. However, smoothing may produce a
biased estimate by displacing activation peaks and underestimating their height.
To address this issue, spatial modeling has been proposed (11,20) to take the spatial
activation pattern into consideration. Recently, spatiotemporal linear regression
methods have also been applied to fMRI data analysis (2,23). These methods use
the time series of neighboring voxels together with their own, and thus take
simultaneously the spatial and temporal correlations into account, which is also
one of the benefits of the regression method to be introduced in this chapter.
18.2.3
MRI D
M
F
ATA
ODELING
In general, techniques for analyzing fMRI data can be divided into model-driven,
e.g., standard general linear model (GLM) (14), and data-driven methods, e.g.,
principal component analysis (PCA) (1), independent component analysis (ICA)
(29), or fuzzy cluster analysis (FCA) (13). In model-driven methods, a model of
the expected response is generated and compared with the data. These methods
require prior knowledge of event timing, from which an anticipated hemodynamic
response can be modeled. However, for brain responses that are not directly
locked to the paradigm, model-driven analysis may not be adequate (8). Data-
driven methods, however, explore the fMRI data statistically without any assump-
tion about the paradigm or the hemodynamic response function. This flexibility
is desirable especially in cases in which it is difficult to generate a good model;
however, there are drawbacks. For example, the assumption implicit in PCA is
that different modes are Gaussian and uncorrelated, whereas ICA assumes that
different modes are non-Gaussian and independent. In addition, a significance
estimate for each component is usually not available. Given the advantages and
disadvantages, a new approach is discussed in this chapter to merge data-driven
methods with prior time course modeling by adjusting a model coefficient.
18.2.4
O
VERVIEW
Despite the progress in fMRI analysis, there is still a need for robust and unified
analysis methods because of the many limitations with existing techniques, as
described in the preceding text. In this chapter, we present a novel, general, and
reliable nonlinear approach for fMRI analysis based on statistical learning method,
i.e., spatiotemporal SVR (ST-SVR), so that existing difficulties resulting from
noise, low resolution, and inappropriate smoothing and modeling can be addressed.
In summary, SVR provides a comprehensive but parsimonious mapping from a
set of input vectors to a scalar output. Its ability to handle highly nonlinear
mappings, in an unconstrained way, makes it a natural candidate for the analysis
Search WWH ::
Custom Search