Classical and Bayesian Inference in fMRI - Advance Image Processing in Magnetic Resonance Imaging - page 529

V4 and V5 as being specialized for the processing of color and motion, respec-
tively. More recently, these analyses have been augmented by functional inte-
gration studies, which describe how functionally specialized areas interact and
how these interactions depend on changes of context. A recent example is the
study by Buchel et al. (2) who found that the success with which a subject
learned an object-location association task was correlated with the coupling
between regions in the dorsal and ventral visual streams (3). In this chapter,
we will address the design and analysis of neuroimaging studies from these
two distinct perspectives but note that they have to be combined for a full
understanding of brain mapping results.
In practice, the general linear model (GLM) is used to identify functionally
specialized brain responses and is the most prevalent approach to characterizing
functional anatomy and disease-related changes. GLMs are fitted to fMRI time
series at each voxel resulting in a set of voxel-specific parameters. These param-
eters are then used to form statistical parametric maps (SPMs) or posterior
probability maps (PPMs) that characterize regionally specific responses to exper-
imental manipulation. ( Figure 17.4 and Figure 17.5 , for example, show PPMs
and SPMs highlighting regions that are sensitive to visual motion stimuli).
Analyses of functional integration are implemented using multivariate
approaches that examine the changes in multiple brain areas induced by experimen-
tal manipulation. Although there are a number of methods for doing this, we focus
on a recent approach called dynamic causal modeling (DCM).
In order to assign an observed response to a particular brain structure or
cortical area, the data must conform to a known anatomical space. Before con-
sidering statistical modeling, this chapter, therefore, deals briefly with how a time
series of images (from single or multiple subjects) are realigned and mapped into
some standard anatomical space (e.g., a stereotactic space).
A central issue in this chapter is the distinction between classical and Bayesian
estimation and inference. Historically, the most popular and successful method
for the analysis of fMRI is SPM. This is based on voxelwise GLM and Gaussian
random field (GRF) theory. More recently, a number of Bayesian estimation and
inference procedures have appeared in the literature. A key reason behind this is
that as our models become more realistic (and, therefore, complex) they need to
be constrained in some way. A simple and principled way of doing this is to use
priors in a Bayesian context. In this chapter we will see Bayesian methods being
used in spatial normalization ( Subsection 17.2.3 ), posterior probability mapping
(Section 17.5) and DCM (Section 17.6). One should not lose sight, however, of
the simplicity of the original SPM procedures (Section 17.4) as they remain
attractive, both from an interpretive and computational perspective.
The analysis of functional neuroimaging data involves many steps that can
be broadly divided into: (1) spatial processing, (2) estimating the parameters
of a statistical model, and (3) making inferences about those parameter esti-
mates with appropriate statistics. This data processing stream is shown in
Figure 17.1 .