Exploratory Data Analysis Methods in Functional MRI - Advance Image Processing in Magnetic Resonance Imaging - page 502

When the parameter m equals 1, we get hard partitioning, and the algorithm
is similar to k -means. When m tends to plus infinity, the membership values tend
to 1/ K . The values usually used in fMRI data analysis range from 1.3, as in
Reference 25, to 2, as in Reference 21. However, there are no precise indications
in the literature for this choice. Another issue is the fact that in fMRI the activated
voxels are a small fraction of the entire population [21], so the clusters found can
be biased by voxels containing artifacts such as movement-related artifacts or
large-vein contributions. This can be resolved by focusing on a region of interest
(ROI) determined with a priori anatomical or functional knowledge or by selecting
the voxels of interest using a first processing stage with statistical tools. In the
study in Reference 19, the fuzzy cluster analysis (FCA) was repeated in a region
of interest, found in a first step, whose centroid showed a correlation with the task.
This second clustering step identified two clusters in the one detected previously:
one contained few voxels with large signal changes and was thought to have
originated from medium-to-large veins running perpendicular to the acquisition
plane and the other contained more voxels with less signal change and was thought
to be related to cortical activation. The problem of orientation of large vessels with
respect to the acquisition plane was pointed out. Another issue is the noise present
in the time series that can affect, as stated before, clustering techniques in general.
Both Toft [26] and Goutte [11], by means of hierarchical and k -means clustering,
tried to overcome this problem by the introduction of a distance based on a function
of the correlation coefficient with a stimulus function, rather than the raw time
series. This approach limited the use of the explorative method to a more selective
exploration and was used also in the study in Reference 12, in which the signifi-
cance of the membership values and the problem of the threshold were addressed.
These values, in fact, do not represent absolute statistical information about the
probability of a time series being correlated to a cluster, but they represent a relative
measure taking into account also the relationship with the other clusters. In the
same work, results were shown with the membership values thresholded at 0.8.
The same value was indicated in Reference 19 as a result of a comparison with a
correlation analysis. Fuzzy clustering for fMRI data has been extensively applied
[12,16-21]. Preprocessing steps can heavily affect the results. Some of the nec-
essary steps have been already mentioned in Section 16.2 and are common pre-
processing techniques. The removal of the baseline level from each time series is
usually applied in order not to allow the clustering algorithm to classify the voxels
based on the underlying anatomical structure. Other preprocessing strategies
involve subtracting the mean value and then dividing by it in order to achieve
a percentage signal change, or the normalization obtained by subtracting the mean
and dividing by the standard deviation. A comparison of different preprocessing
strategies can be found in Reference 12.

16.3.2.4 Artificial Neural Networks

Artificial neural networks have been extensively used for clustering and classifi-
cation purposes. In particular, competitive neural networks are useful because