Image Processing Reference
In-Depth Information
refined techniques from statistics, biosignal analysis, and image processing and
analysis is required for sensitive and robust detection and characterization of func-
tional activity.
This chapter is organized as follows: Section 18.2 provides some background
about fMRI and its data analysis. Section 18.3 gives the concepts and theory of the
statistical learning methods, support vector machines (SVMs) and support vector
regression (SVR). The proposed framework and its features are introduced and
described in detail in Section 18.4, with results on both simulated and real fMRI
data. Section 18.5 concludes the chapter with further discussions on this work.
18.2
BACKGROUND
18.2.1
N
MRI
ONLINEARITIES
IN
F
The BOLD signal is a complex function of neural activity, oxygen metabolism,
cerebral blood volume, cerebral blood flow (CBF), and other physiological param-
eters. The dynamics underlying neural activity and hemodynamic physiology are
believed to be nonlinear (3,12,16). The observed fMRI response to a stimulus
consists of two chain reactions: The stimulus first triggers a neural response,
which sequentially triggers a hemodynamic response that is recorded by BOLD
fMRI. The nonlinearity in fMRI could arise from either a nonlinearity in the
neural response or a nonlinearity in the hemodynamics, or both (4,30,41). The
spatial heterogeneity of the nonlinear characteristics of BOLD signals has also
been reported in the literature (3,21).
The cascade of neuronal and hemodynamic nonlinearities in the system would
make the determination of variations in neuronal activity difficult. For simplicity,
most existing fMRI data analyses assume a linear convolution model and primarily
rely on linear methods or general linear models (GLMs). However, as fMRI exper-
iments have grown more sophisticated, the role of nonlinearities is becoming more
important under certain situations. Some authors have investigated the physiological
mechanisms that reveal the relationship between synaptic activation and vascular
or metabolic controlling systems (22,27). Accordingly, initial attempts that model
the BOLD signal at macroscopic levels have been made by using differential
equations, linking the hemodynamical variations with physiological sense (7,16,34).
Although these theoretical models have high impact on fMRI analysis, solid valida-
tion from real data is still needed in order to justify their practical use. Because of
the complexity of the human brain, we propose to approach the whole brain through
a more flexible and general model, and filter the noisy fMRI signals using a nonlinear
statistical learning method, support vector regression (SVR). The restored signals
are considered nonlinear functions or responses of the stimulus reference function,
which agrees with recent findings about the presence of nonlinearities in fMRI.
18.2.2
F
MRI D
EATURES
OF
F
ATA
Two features peculiar to fMRI make its analysis more challenging. First, fMRI
data have intrinsic spatial and temporal correlations (42). Second, fMRI data tend
Search WWH ::




Custom Search