Exploratory Data Analysis Methods in Functional MRI - Advance Image Processing in Magnetic Resonance Imaging - page 495

Xi (3)

Xi (t) = {Xi(1), Xi (2), Xi(3)}

Xi (2)

Xi (1)

FIGURE 16.2 A 3-point-long time series, depicted as a vector in R

3

. The same procedure

can be extended to n -dimensional time series.

correct for movement (see Chapter 17 ). Spatial filtering in order to increase

contrast-to-noise ratio (CNR) or temporal filtering in order to remove low-
frequency fluctuations can be used: low-frequency drifts due to slow subject
movements or instrumentation changes can be removed by means of linear regres-
sion detrending. In general linear model (GLM) approaches [6,7], spatial filtering
is usually performed by means of an isotropic spatial filter that ignores the
anatomical structure of the brain. This operation can be required in order to satisfy
the hypothesis of spatial smoothness of the activated regions as in Gaussian
random field theory [47] and are used in the successive inferential steps. In the
case of exploratory methods this is not required and different approaches can be
used in order to preserve anatomical information, such as nonlinear filtering, for
example, the SUSAN filter [48], or spatiotemporal approaches, for example, those
in Reference 49.

16.3 DATA CLUSTERING APPROACHES

Clustering methods try to group data set elements following a similarity criterion:
the elements belonging to a cluster should be similar. The clustering procedure
results in a classification of the data set in order to distinguish between different
signal sources, both physiological, such as task-related activations, and artifactual,
such as movement-related effects. Clustering techniques are suitable for fMRI
data analysis because interesting phenomena, such as task-related activation, will
involve several voxels, which can be grouped together without any

knowl-
edge about the shape or the extension of the activations. Experimenters who want
to use clustering techniques have to be aware that there are many possible choices
for the algorithm and the preprocessing steps. Because the clustering techniques

a priori