7. Results

Because of the large amount of subjects and since showing the volumetric activation patterns of the independent components on paper can be problematic, this chapter highlights only the most interesting and surprising results. The complete results for all individual experiments are shown in Appendix B. Although similar components are found on many subjects, as discussed in Section 7.2, the results are shown using the clearest examples. The visualization is explained in detail in Chapter 5 and the information on the human brain, needed to understand the medical explanations, is in Chapter 2.

7.1.1 Overview

Overall, the individual results shown in Figures B.1 - B.14 are very good, and clearly demonstrate the usability of the analysis method and visualization tools. They are easy and reliable to interpret. The results show that the components closely related to the stimulus, explained in Chapter 6, are very consistent. There are also many other consistent components that have either an interesting spatial or temporal structure, or both. Additionally, some of the less consistent components are still interesting, perhaps revealing surprising phenomena.

Particularly, the results for subjects JK, PK and SN seem to contain many interesting components related to the stimulus. The majority of those components are also highly consistent. On the other hand, there seems to be a very strong scanning anomaly in the data of subject TL. Also, a similar phenomenon can be seen in the results for subject MG, but in that case the artifact appears to contaminate the whole results.

Note how the measures of discrimination, variability and skewness of a component together seem to reveal how reliable, and possibly interesting, the component is. The results are interesting both from a theoretical point of view, validating the method, and from a medical point of view, possibly revealing unexpected relations within human brain processing of speech and language.

7.1.2 Components Related to Stimulus

Components that are closely related to the stimulus are quite predictable and should be very consistent with all analysis methods. For example, Figure 7.1 shows two interesting components, the first from subject PK and the second from subject SN. Similar components can be found practically from all subjects, but these are shown just as two very clear cases. The first component appears to correspond to activation on the primary auditory areas, and mainly on the right hemisphere. The second component seems to relate to a secondary auditory area, sometimes referred to as Wernicke's area. The activation on the second component is clearly stronger on the left-hand side. The figure shows only the three slices centered on the point of strongest activation, but even without the possibility of interactively studying the whole volumes, the skewed histograms support the fact that the volumes contain very strong and focal activation.

Both components are also very consistent with time-courses strongly related to the stimulus, with the first component being almost a perfect match. This is predictable, since the activation on auditory areas should follow the speech stimulus quite closely. Additionally, the discrimination of the components is very good, that is, they are well separated from other components. The disks depicting the intra-group distances are mere dots and would easily fit inside the inter-group disks, although the inter-group distances have a somewhat bigger spread on the second case.

**Figure 7.1:** Closer look at consistent stimulus related components. The (a) primary auditory areas and (b) secondary auditory areas.
$\includegraphics[width=0.48\textwidth]{images/results_primaryauditory.eps}$ (a)	$\includegraphics[width=0.48\textwidth]{images/results_additionalauditory.eps}$ (b)

7.1.3 Components Revealing Artifacts

Two examples of clear artifacts are shown in Figure 7.2. The first is from subject JK and related to filtering, and the second reveals a very strong scanning anomaly in the first scan from subject TL. Both components are well discriminated and very consistent. Considering the complete results for these subjects, see Figures B.4 and B.11 respectively, it is clear that ICA is able to remove the artifacts from the rest of the data by isolating them as separate components. Although the multiple runs of ICA, explained in Chapter 4, are able to separate the filtering related artifact, the spread of inter-group distances is still quite large. Clearly, the artifact that affects only a single time point is easier to separate.

The histogram of the first example is very symmetric, suggesting that there should not be any strong focal activation in that volume. Actually, a closer look at the whole volume reveals it to be completely covered by a pattern that resembles ripples, like often seen on the surface of water. The fast fMRI scanners often produce such artifacts, and the signals are actually filtered to remove such patters. However, it seems that the phenomenon causing the artifact is not fully stationary, and the simple constant filtering still leaves part of the contamination with an almost linearly drifting time-course. This new information, revealed by ICA, could be used to improve the filters used in the scanning.

On the other hand, the histogram of the second artifact is very skewed and the time-course shows that the artifact is only present in the first volume of the fMRI sequence. This artifact seems to result from a strong ghosting or shadowing phenomenon located just below the frontal lobe of the brain. It is almost completely outside the volume used in the analysis, and only partially visible in the example slices.

**Figure 7.2:** Closer look at clear aftifact components. (a) A filtering artifact and (b) a scanning anomaly in the first volume of the sequence.
$\includegraphics[width=0.48\textwidth]{images/results_lineardrift.eps}$ (a)	$\includegraphics[width=0.48\textwidth]{images/results_scanninganomaly.eps}$ (b)

7.1.4 Components with Strong Variability

The examples in Figure 7.3 show the true benefits of the multiple run method. These cases are more complex than the previous ones, and would practically be impossible to analyze reliably using only a single run of ICA, or the traditional SPM method explained in Section 2.3.3. Both examples show a component with strong variability, yet, the nature of the variability is very different in each case.

The component in Figure 7.3(a), from subject KR, seems to contain activation on the visual area. The activation pattern itself seems rather clear with a relatively skewed histogram, much like the auditory examples in Figure 7.1. Additionally, activation on the visual cortex is common in all kinds of studies, since the eyes provide a constant flow of information, even when the subject would concentrate on listening. However, the strong temporal variability suggest that the component is not very consistent, or reliable. Indeed, focusing on the disks showing the discrimination measures reveals that the intra-group distances are very small, excluding an outlier, but the inter-group distances are more uniformly spread and the discrimination is not very good. This means that the component estimation itself may be stable, but the component is very difficult to separate from other components. One explanation for such behavior could be that the activation on the visual area may be too weak, or unstructured, to detect reliably under the auditory stimulation.

On the other hand, the situation in Figure 7.3(b), which shows a component from subject SN, is very different. The highly symmetric histogram suggests that the activation pattern does not contain very strong points. Still, there seems to be some focal patterns around the brain stem. In this case, the small spread and large mean of the inter-group distances suggest that the component is quite well separated from other components. The estimation of the component itself is much more unstable than that in the previous case, since the mean value of the intra-group distances is quite big, and there is actually a hole in the middle of the disk. Also, the very high number of grouped estimates supports the interpretation that the component is able to identify a distinct, but very unstable, phenomenon. Most probably, the component reveals a portion of the main blood vessels, rising around the brain stem. Since the fMRI scanning is too slow to accurately follow the flow of oxygenated blood in the vessels, ICA could detect a separate signal subspace, but seems to be unable to accurately decompose that subspace into independent components. Additionally, the time-course of the component is very structured, although highly varied, and could be related to the heart beat.

**Figure 7.3:** Closer look at components with strong variability. (a) Activation on the visual area and (b) a possible portion of the blood vessels.
$\includegraphics[width=0.48\textwidth]{images/results_visual.eps}$ (a)	$\includegraphics[width=0.48\textwidth]{images/results_bloodvessel.eps}$ (b)

7.1.5 Other Interesting Components

Two other interesting examples are shown in Figure 7.4. The first from subject PK, showing potential activation in the Broca's area, and the other, from subject JK, which reveals activation around the frontal cingulate gyrus. The components are very consistent, and the time-courses suggest that they are not completely unrelated to the stimulus. These components are interesting because Broca's area is involved in the production of sentences and areas around the cingulate gyrus may relate to long-term memory. Activation around these areas while listening to spoken language is quite reasonable. Still, the components are so weakly related to the stimulation that they would be very difficult to detect using the traditional SPM method.

**Figure 7.4:** Closer look at some of the other interesting components. Activation possibly on (a) the Broca's area and (b) frontal cingulate gyrus.
$\includegraphics[width=0.48\textwidth]{images/results_broca.eps}$ (a)	$\includegraphics[width=0.48\textwidth]{images/results_cingulate.eps}$ (b)

7.1.6 Spatial Relations Revealed by Variability

As mentioned in Chapter 4, the variability of a component can also be analyzed spatially by calculating the spread of the group using the volumes. Although the computation using the whole volumes is quite demanding, the resulting volumes of spatial variability can be used just like the temporal spread around the mean time-courses.

A very interesting example is shown in Figure 7.5, from subject SN, where the spatial variability reveals a relation that can not be seen from the time-courses alone. A slice of component 10 is shown with that of component 7. The second slice of component 10, shown in the middle, is the spatial variability. The spatial pattern of component 7 is remarkably similar to the variability of component 10. This suggests that the main source of variability for component 10 is actually component 7, that is, the estimation of component 10 tends to, from time to time, get mixed with component 7. The components may, for example, span a subspace together.

**Figure 7.5:** Spatial variability clearly reveals related components from subject SN. Horizontal slices of (a) component 10 and (b) its spatial spread next to (c) component 7. The main source of variability for component 10 is clearly component 7.
$\includegraphics[width=0.2\textwidth]{images/results_link_ic10.eps}$ (a) IC #10	$\includegraphics[width=0.2\textwidth]{images/results_link_var10.eps}$ (b) VAR #10	$\includegraphics[width=0.2\textwidth]{images/results_link_ic7.eps}$ (c) IC #7

7.2 Group Results

Although the scope of the thesis does not include group analysis, the amount of information provided by the method allows quite fast and reliable human comparison of the subjects. The similarities, and differences, among the subjects are easy to spot with the help of the variabilities. Some of the components, shown only as horizontal slices in Appendix B, may be quite hard to identify without the possibility to interactively study the whole volumes. On the other hand, the task is not simple as similarities can be identified in many ways. Subjects may share a spatially matching activation pattern with possible differences in the time-courses or the time-courses may be similar with some differences in the spatial patterns. In some cases, both may be very similar.

Some of the components shared by multiple subjects are summarized in Table 7.1. The components used in the table are (PA) the primary auditory areas, as in Figure 7.1(a); (BV) the blood vessels around the brain stem in Figure 7.3(b); (AA) secondary auditory areas, as in Figure 7.1(b); (C) the areas around the cingulate gyrus, as in Figure 7.4(b); (VE) large ventricles located in the middle of the brain; (VI) the visual area in Figure 7.3(a); and (F) the filtering artifact in Figure 7.2(a).

The comparison suggests that the subjects can be categorized into three groups. In the first, the majority of the subjects share many of the components Second, some subjects seem to have only a subset of the components, which may be caused by, for example, individual differences or level of attention. And third, a few subjects share almost no similarities with the others. This may be explained by the fact that the data of these few subjects seems to be very heavily contaminated by scanning artifacts.

Table 7.1: Consistency across subjects shown as a table of very similar components shared by multiple subjects. Definitions of the components: PA is the primary auditory areas, BV is the blood vessels around the brain stem, AA is secondary auditory areas, C is the cingulate gyrus, VE is the main ventricles, VI is the visual area and F is the filtering artifact.

		Components
		PA	BV	AA	C	VE	VI	F
SN	B.10	$\bullet$	$\bullet$	$\bullet$	$\bullet$	$\bullet$	$\bullet$	$\bullet$
KR	B.5	$\bullet$	$\bullet$	$\bullet$	$\bullet$	$\bullet$	$\bullet$	$\bullet$
TL	B.11	$\bullet$	$\bullet$	$\bullet$	$\bullet$	$\bullet$	$\circ$	$\bullet$
MT	B.7	$\bullet$	$\bullet$	$\bullet$	$\bullet$	$\bullet$	$\bullet$	$\circ$
JK	B.4	$\bullet$	$\bullet$	$\bullet$	$\bullet$	$\circ$	$\circ$	$\bullet$
RS	B.9	$\bullet$	$\bullet$	$\bullet$	$\bullet$	$\circ$	$\circ$	$\bullet$
TT	B.13	$\bullet$	$\bullet$	$\bullet$	$\bullet$	$\circ$	$\bullet$	$\circ$
HH	B.2	$\bullet$	$\bullet$	$\bullet$	$\circ$	$\bullet$	$\bullet$	$\circ$
HR	B.3	$\bullet$	$\bullet$	$\bullet$	$\circ$	$\bullet$	$\bullet$	$\circ$
PK	B.8	$\bullet$	$\bullet$	$\bullet$	$\circ$	$\bullet$	$\circ$	$\circ$
AS	B.1	$\bullet$	$\circ$	$\circ$	$\bullet$	$\circ$	$\bullet$	$\bullet$
UL	B.14	$\bullet$	$\bullet$	$\circ$	$\circ$	$\circ$	$\circ$	$\bullet$
TP	B.12	$\bullet$	$\bullet$	$\circ$	$\bullet$	$\circ$	$\circ$	$\circ$
MG	B.6	$\bullet$	$\circ$	$\circ$	$\circ$	$\bullet$	$\circ$	$\circ$