Functional Connectivity - a brief overview

Background
During its initial application in scientific research, fMRI was used primarily with task-based behavioral studies. It was used to locate brain activation associated with a certain task or stimulus. For example, we want to study V1 in response to seeing patterns of a checkerboard, or S1 in response to a tactile stimulus on our left forearm. Since Biswal et al., (1995) found a temporal correlation in the sensorimotor area in the absence of tasks, the interest in the so-called spontaneous fluctuations in the brain has accelerated. This type of fMRI, the resting-state fMRI, is unique because the frequency content is low < 0.1 Hz. That is why it was initially thought of as mere physiological noise.

In another study, Fox et al. (2005) reveal a set of interesting brain regions where they get deactivated after a person engages in a task. These regions are known to routinely exhibit task-negative responses (deactivations) during attention-demanding tasks. For terminology sake, brain regions that get deactivated during task performance are called task-positive. When talking about spontaneous fluctuations, we have the terms correlation and anticorrelation regions. Anticorrelation signifies a form of temporal negative correlation. In that study, the fluctuation that is negatively correlated with those in the attention and executive systems is later called the default mode network.

Statistical maps produced by this spontaneous BOLD acquisition is thought to reflect functional connectivity network or resting-state network (RSN). Some authors (Seekey, 2007) used the word intrinsic connectivity to denote two distinct and dissociable networks, the "salience network," (dACC and orbital frontoinsular cortices, with connectivity to subcortical and limbic structures), and "executive-control network" that links dorsolateral frontal (dPFC) and parietal neocortices of the brain. These intrinsic networks are correlated with task-based MRI measured.

Fig-1: (A) BOLD activity during finger tapping and spontaneous fluctuation with a seed located at point a. (B) Spontaneous deactivation observed during task performance using PET. Regions that are negatively correlated with the seed a is called the default mode network.

Methodological Perspective
As mentioned, the only difference between rs-fMRI and task-based fMRI is that the acquisition is performed at rest. The subjects typically lie down in the scanner with their eyes open and looking at a cross-hair, or with eyes closed. Some differences in brain activations exist between the two variants. The minimal requirements of an rs-fMRI study make it easy to scan a wide variety of populations, subjects may get bored and fall asleep. A recent study indicates that falling asleep is the major challenge for subjects and that fixating subjects rarely fall asleep.

The way we analyze resting-state fMRI data is different from the method used in the task-based fMRI. This is because we do not have any clear task model for our design matrix. Having said that, some steps in the preprocessing pipeline are still valid here. For example, the removal of non-brain tissues and skulls, spatial smoothing, high pass filtering, image registration, and motion correction.

Resting-state fMRI suffers two essential drawbacks. First, separating wanted signals from the image artifacts produced by magnetic field distortion (e.g. near the sinuses), signal dropout, and motion can be challenging. Unless we can correct or model these events, the presence of artifacts potentially reduces the temporal correlation between two brain areas. Image distortion due to sinuses can be corrected using field map correction. FSL/FreeSurfer software package has a subroutine called BBR that helps to correct this when registering the EPI to T1-image.

Second, physiological signals such as cardiac and respiratory cycles can also modulate the wanted signals via several mechanisms (Murphy et al., 2013). More recent works have shown that, with a very high sampling rate or TR < 400 msec, we can segregate the frequency contents of spontaneous brain signals. But such MRI sequences are still work-in-progress. There are, however, a few methods to correct for this:

• Regress out the average time series obtained within the white matter region and ventricles, which is mostly of no interest.
• Regress out respiratory and heart-rate signals acquired from the actual physiological recording.
• Regress out global (whole-brain) average time series. This method is still debatable as it allegedly introduces the systematic anticorrelation areas.
• Use ICA to identify artifactual components around the perimeter, cardiac-related, or anything with the frequency > 0.1 Hz.
• Use band-pass filtering to limit the frequency content to be 0.01 - 0.1 Hz.

Once we have more or less 'clean up' our dataset, there are two most common and essential methods in resting-state data analysis, i.e. seed-based and ICA (Independent Component Analysis) methods. The methodological aspects of rs-fMRI are massive and it is always evolving. But some key points I'd like to write about RSN analyses:

• The frequency spectrum is usually between 0.01 - 0.11 Hz, i.e. the 'low frequency' spontaneous fluctuation, although a wider range up to 0.25 Hz was recently accepted.
• There are two most commonly used methods to determine RSN: the seed-based correlation analysis (SCA) and Independent Component Analysis (ICA).
• The SCA method is a model-driven analysis that requires a strong apriori region of interest selection. This method allows us to use the usual framework of general linear modeling (GLM). Example: if we are studying the effect of playing tennis on the brain, we should select M1 as one of our seeds.
• ICA is a data-driven analysis and does not require apriori assumption. ICA is also very attractive in identifying various neural activities at rest, collectively reflected as independent components, e.g. physiological artifacts and jerky head movements.
• It is important that changes in RSN do not constitute the causality of our behavioral paradigm.

Exploring Brain Organization
In several studies, ICA has been used as a robust method to segregate different functional networks of the brain (RSNs). With improvements in the MRI sequences, more and more details can be picked up using rs-fMRI techniques. The number of wanted components was found to be 10 of 25 in (Damoiseaux et al., 2006), 45 of 70 in (Smith et al., 2009), and ~23 of ~200 in (Marcus et al., 2013). Another method is more like a seed-based approach, which involves calculating pairwise correlations between all voxels and using clustering algorithms to identify groups of highly correlated voxels. Those studies identified the brain as being partitioned into somewhere around 15–20 large-scale clusters (e.g. Power et al., 2011; Yeo et al., 2011). Refer to Fig-2.

Fig-2: Various resting state networks revealing different functional areas of the brain (motor, sensory, visual etc).

Some of these networks have been known anatomically for long ago but others, such as attentional networks, are new. Interestingly, in some cases, functional connectivity grouped brain areas that were not so widely recognized as connected. With sophisticated data-driven approaches, research has been spent to refine boundaries between adjacent functional brain regions.

... Resting-state correlations grouped a set of regions and made it easier to recognize that they shared a variety of specific characteristics, bolstering the case that these regions form a functional system. The large-scale (system-level) patterns in resting-state activity, therefore serve as a useful organizing framework for interpreting results and patterns in other modalities.

Recently, a new paradigm of analyzing rs-fMRI data called the network-based approach emerges. When viewing the data as a network, less focus is given on the properties of a single brain area, but more within the larger neural system. However, it is wise to note that a network in this sense is not based on anatomical or physical neuronal networks. Instead, it is from time-varying signals arising of either BOLD (functional networks) or DTI (diffusion), which are essentially mathematical models with some underlying assumptions, that may not fully reflect the physical entity.


Source: Power, JD, Schlaggar BL, Petersen, SE (2014). "Studying Brain Organization via Spontaneous fMRI Signal", Neuron-Primer. Some other notable papers: (Raichle, 2010) for a historical and metabolic perspective; (Deco et al., 2011; Hutchison et al., 2013) for dynamical perspectives; (Bullmore and Sporns, 2012; Sporns, 2014) for network perspectives; (Murphy et al., 2013) for a methods perspective; (Lee et al., 2013) for a clinical perspective; and (Buckner et al., 2013; Craddock et al., 2013) for general perspectives.