voxels. Additionally, the volumes were masked with a
cortical mask to remove uninteresting voxels outside the brain. This
effectively lowered the amount of data to half, still leaving an intimidating
observation matrix per subject.
To test the usefulness of the ICA analysis and visualization method, explained in Chapter 4, the data from a real functional magnetic resonance imaging (fMRI) study was analyzed. It is important that the method is tested with real world data, since an artificial data set would not prove the usefulness under realistic noise and variability structures. Additionally, the method might reveal unseen phenomena from the data.
The experiments involved 14 voluntary subjects, which are referred to only by their initials to protect their anonymity. The stimulus and scanning conditions were repeated as closely as possible for each subject to allow the comparison of results between subjects.
The study used an auditory stimulus, which consisted of repetitions of spoken text with resting periods in between. These periods were repeated four times during the experiment. The results should reflect this by showing activation at least on the primary auditory areas, but also on additional language and memory related areas. The activation time-courses of these areas should be somewhat related to the stimulus, especially on the primary areas.
During the four repetitions of speech and resting periods 10 full head volumes were acquired in each condition with a scanning interval of approximately 3 seconds, resulting in a total of 80 volumes. The scanning was done using a 3 Tesla GE scanner in the Advanced Magnetic Imaging Centre (AMI-Centre) of the Helsinki University of Technology.
It is common that under a hypothesis driven experiment the scanning is focused only on a few slices of the brain, which have been classified as interesting beforehand. Sometimes this is beneficial, since it could allow faster scanning or increased resolution, but as ICA is a purely data-driven method it would be rather impossible to define the interesting regions of the brain beforehand. And again, such scanning would limit the data too much to fully characterize unexpected phenomena.
Before analysis, the volumes were processed in the usual way
(c.f., Worsley and Friston, 1995) fMRI data is processed when using
the traditional analysis method (see Section 2.3.2).
This was done with the SPM toolbox (SPM, 1999) and resulted, for each
of the 14 subjects, in 80 volumes with a resolution of
voxels. Additionally, the volumes were masked with a
cortical mask to remove uninteresting voxels outside the brain. This
effectively lowered the amount of data to half, still leaving an intimidating
observation matrix per subject.
The observation matrix of each subject was then analyzed with the method described in Section 4.3. The FastICA algorithm was run 100 times using a bootstrapping of 20% and PCA whitening to the 30 strongest principal components in each run. FastICA was used in symmetric mode with nonlinearity tanh, estimating 15 independent components in each run.
This resulted in 1500 independent component estimates per subject, which were then clustered with correlation threshold 0.8 and power 8. All parameter values in the experiments were chosen heuristically, based on earlier trials.
As mentioned earlier the described individual experiment was repeated under the same conditions for all 14 subjects. Although the analysis of consistency within the whole group is beyond the scope of this thesis, some general observations can still be possible and could prove valuable. Therefore the complete results for all subjects were kept in an easily comparable form.