Exploratory Data Analysis Methods for fMRI - Biomedical Signal Analysis: Contemporary Methods and Applications - page 255

8

Exploratory Data Analysis Methods for fMRI

Functional magnetic resonance imaging (fMRI) has been shown to be
an effective imaging technique in human brain research [188]. By blood
oxygen level- dependent contrast (BOLD), local changes in the magnetic
field are coupled to activity in brain areas. These magnetic changes are
measured using MRI. The high spatial and temporal resolution of fMRI
combined with its noninvasive nature makes it an important tool for
discovering functional areas in the human brain and their interactions.
However, its low signal-to-noise ratio and the high number of activities
in the passive brain require a sophisticated analysis method. These
methods either (1) are based on models and regression, but require
prior knowledge of the time course of the activations, or (2) employ
model-free approaches such as BSS by separating the recorded activation
into different classes according to statistical specifications without prior
knowledge of the activation.
The blind approach (2) was first studied by McKeown et al. [169].
According to the principle of functional organization of the brain, they
suggested that the multifocal brain areas activated by performance of
a visual task should be unrelated to the brain areas whose signals
are affected by artifacts of a physiological nature, head movements, or
scanner noise related to fMRI experiments. Every single process can
be described by one or more spatially independent components, each
associated with a single time course of a voxel and a component map.
It is assumed that the component maps, each described by a spatial
distribution of fixed values, represent overlapping, multifocal brain areas
of statistically independent fMRI signals. This is visualized in figure 8.1.
In addition, McKeown et al. [169] considered the distributions of the
component maps to be spatially independent and in this sense uniquely
specified (see section 4.2). They showed that these maps are independent
if the active voxels in the maps are sparse and mostly nonoverlapping.
Additionally, they assumed that the observed fMRI signals are the
superpositions of the individual component processes at each voxel.
Based on these assumptions, ICA can be applied to fMRI time series
to spatially localize and temporally characterize the sources of BOLD
activation. Considerable research has been devoted to this area since the
late 1990s.