**2. Physiology and anatomy of the resting-state networks**

In a functional connectivity study, the active human brain is conveniently represented in terms of independent functional networks of mutually interacting regions (Friston et al., 1996). Anatomically, these regions can be either distant from each other or result from a fine segregation of bigger into smaller neuronal assemblies. Functionally, two imaging voxels or regions exhibiting neural signals highly correlated in time (synchronized) are conceptually part of the same network.

When measuring the brain with fMRI over prolonged intervals of time (e. g. from two to ten minutes), it is possible to detect throughout the brain characteristic spontaneous fluctuations in the BOLD signal which occur at relatively low frequencies (<0.1 Hz) and are not of technical (artifactual) origin (Biswal et al., 1995; Damoiseaux et al., 2006; Mantini et al., 2007; Smith et al., 1999; van de Ven et al., 2004). When assessing the functional connectivity of these fluctuations across regions, a series of networks can be built, that clearly resemble the same functional networks activated during the performance of active tasks, even if the participant is not performing any specific task and is simply instructed to remain still, with eyes closed and without thinking to anything specific. In other words, under simple resting conditions, the brain is engaged in spontaneous activity which is not attributable to specific inputs or to the generation of specific output, but is intrinsically originated.

A possible link between functional connectivity and spontaneous synchrony of brain signals was already proposed in early electroencephalography (EEG) studies (French & Beaumont, 1984), suggesting that long-range EEG coherencies across cortical regions and between hemispheres could be originated from a relatively small number of interacting regions and processes (see, also, e.g., (Koenig et al., 2005; Locatelli et al., 1998)). Even if EEG and fMRI signals have very different spatial and temporal scales, it has been later demonstrated that a tight correspondence exists between the spatial distribution of low-frequency BOLD signal fluctuations and the main EEG rhythms, for at least six reproducible RSNs (Mantini et al., 2007), in line with the idea that the two modalities share a common neurophysiological origin, represented by the local field potential (LFP) (Logothetis & Pfeuffer, 2004). More specifically, it has been hypothesized that the low-frequency BOLD signal fluctuations are themselves due to low-frequency LFPs or to low frequency modulations of high-frequency LFPs (Raichle, 2010).

The most frequently and consistently reported RSNs, which are also correlated with EEG rhythms, include the default-mode network (DMN), functionally connecting the posterior and anterior cingulate cortex (PCC and ACC) and, bilaterally, the inferior parietal lobules

primary or secondary aspect of neurodegeneration (Bonavita et al., 2011; Cherkassky et al., 2006; Greicius et al., 2007; Greicius et al., 2004; Mohammadi et al., 2009; Nakamura et al., 2009; Rocca et al., 2010; Rombouts et al., 2005; Roosendaal et al., 2010; Sorg et al., 2007; Sorg

In this chapter we will review the physiological and technical background of resting state neural networks and the ICA methodology currently used for observing and analyzing RSNs in normal and clinical populations. The main physiological RSNs will be illustrated and discussed with special emphasis to those exhibiting functional abnormalities in neurological disorders. In addition, two clinical applications will be presented, where this methodology showed pathological changes in amyotrophic lateral sclerosis (ALS) and

In a functional connectivity study, the active human brain is conveniently represented in terms of independent functional networks of mutually interacting regions (Friston et al., 1996). Anatomically, these regions can be either distant from each other or result from a fine segregation of bigger into smaller neuronal assemblies. Functionally, two imaging voxels or regions exhibiting neural signals highly correlated in time (synchronized) are conceptually

When measuring the brain with fMRI over prolonged intervals of time (e. g. from two to ten minutes), it is possible to detect throughout the brain characteristic spontaneous fluctuations in the BOLD signal which occur at relatively low frequencies (<0.1 Hz) and are not of technical (artifactual) origin (Biswal et al., 1995; Damoiseaux et al., 2006; Mantini et al., 2007; Smith et al., 1999; van de Ven et al., 2004). When assessing the functional connectivity of these fluctuations across regions, a series of networks can be built, that clearly resemble the same functional networks activated during the performance of active tasks, even if the participant is not performing any specific task and is simply instructed to remain still, with eyes closed and without thinking to anything specific. In other words, under simple resting conditions, the brain is engaged in spontaneous activity which is not attributable to specific

A possible link between functional connectivity and spontaneous synchrony of brain signals was already proposed in early electroencephalography (EEG) studies (French & Beaumont, 1984), suggesting that long-range EEG coherencies across cortical regions and between hemispheres could be originated from a relatively small number of interacting regions and processes (see, also, e.g., (Koenig et al., 2005; Locatelli et al., 1998)). Even if EEG and fMRI signals have very different spatial and temporal scales, it has been later demonstrated that a tight correspondence exists between the spatial distribution of low-frequency BOLD signal fluctuations and the main EEG rhythms, for at least six reproducible RSNs (Mantini et al., 2007), in line with the idea that the two modalities share a common neurophysiological origin, represented by the local field potential (LFP) (Logothetis & Pfeuffer, 2004). More specifically, it has been hypothesized that the low-frequency BOLD signal fluctuations are themselves due to low-frequency LFPs or to low frequency modulations of high-frequency

The most frequently and consistently reported RSNs, which are also correlated with EEG rhythms, include the default-mode network (DMN), functionally connecting the posterior and anterior cingulate cortex (PCC and ACC) and, bilaterally, the inferior parietal lobules

multiple sclerosis (MS) patients in comparison to normal subjects.

**2. Physiology and anatomy of the resting-state networks** 

inputs or to the generation of specific output, but is intrinsically originated.

et al., 2009; Tedeschi et al., 2010).

part of the same network.

LFPs (Raichle, 2010).

(Greicius et al., 2003; Raichle et al., 2001); the visual network (VIS) involving bilaterally the retinotopic occipital cortex up to the temporal-occipital junctions and middle temporal gyri (Lowe et al., 1998; Wang et al., 2008); the fronto-parietal network (FPN) including, bilaterally, the intra-parietal cortex and the superior-lateral frontal cortex (Corbetta & Shulman, 2002); the sensori-motor network (SMN) involving, bilaterally, the pre- and postcentral gyri, the medial frontal gyrus, the primary and supplementary motor and the primary and secondary sensory areas (Biswal et al., 1995); the auditory network (AUD), involving, bilaterally, the superior and middle temporal cortex (Seifritz et al., 2002) and the self-referential network (SRN) involving the ventro-medial prefrontal cortex and the perigenual anterior cingulate cortex (D'Argembeau et al., 2007). The brain maps of these six typical RSNs in a normal population are exemplarily shown in figure 1 using different colors for the different networks.

Fig. 1. Typical RSN maps. Visual network (VIS), default-mode network (DMN), frontoparietal network (FPN), sensori-motor network (SMN), auditory network (AUD) selfreferential (SRN) network.

As anticipated (and exemplified in figure 1), all the RSNs consist of anatomically separated, but functionally connected regions, sharing and supporting the same sensitive, motor or cognitive functions (Cordes et al., 2000). The RSNs reported in the normative literature have generally resulted to be quite consistent across studies, despite some differences in data acquisition and analysis techniques that partially account for the variability observed in the number and lay out of the networks. For instance, the DMN has been sometimes distinguished into two separate subnetworks, the anterior and posterior DMN (see, e. g., (Damoiseaux et al., 2008)), and the FPN as two lateralized networks (right and left FPNs, RFPN and LFPN) (see, e. g., (Damoiseaux et al., 2006; Tedeschi et al., 2010)). The auditory and visual networks have been presented in terms of a one single network (see, e. g., (Mantini et al., 2007)), two (see, e. g., (Damoiseaux et al., 2006)) or even three (see, e. g., (Rocca et al., 2011)) subnetworks.

Understanding the functional correlate of a given RSN under normal physiological conditions is crucial to correctly address any possible link between altered rs-fMRI patterns and behavioral and clinical variables. However, it should also be recognized that cytoarchitectonically distinct brain regions are kept functionally connected by white matter

Neuronal Networks Observed with Resting State

data sets, thereby avoiding the possibility of bias.

terms of whole-brain distributed maps (Mantini et al., 2007).

al., 2009).

Functional Magnetic Resonance Imaging in Clinical Populations 113

defined region-of-interest (ROI) and subsequently searching all regions whose time-course significantly correlates with the ROI time-course. This method produces RSN maps that are extremely simple to interpret (Fox & Raichle, 2007; Greicius et al., 2003), but has the important drawback that the resulting functional connectivity maps will depend on the location, extension and order of the "seed" regions chosen, and on how these are defined in advance of the analysis. By contrast, "component-based" statistical techniques (Andersen et al., 1999; Friston et al., 1993), that do not require a-priori assumptions on the regions involved, enable the observation of multiple neural networks from whole-brain resting state

ICA (Hyvarinen et al., 2001) has been successfully applied to neuroimaging data of diverse imaging modalities for generating convenient representations of activated brain networks in single subjects and groups. Particularly, in fMRI, ICA is commonly applied in its spatial variant (Calhoun et al., 2001b; McKeown et al., 1998) where each statistically independent component process corresponds to a spatial map distributed over all voxels of the imaging slab. Besides separating many types of structured dynamic artefacts from fMRI time series (see, e. g., (De Martino et al., 2007)), spatial ICA can provide a meaningful representation of function-related BOLD signals and unravel the whole-brain distributed functional connectivity under different experimental and clinical conditions. Particularly, spatial ICA is commonly applied in rs-fMRI to model the spontaneous low-frequency BOLD signals in

When exploring fMRI data with spatial ICA, it is always necessary to decide how many ICA components to extract and, among these, select those components that can be consistently and reliably associated with functional connectivity networks of interest for a given application. The number of components is basically a "free choice" parameter (Calhoun et al., 2009), typically ranging between 20 and 60, even if potential changes in the layout of certain ICA generated RSN maps, such as splitting of a network into multiple networks, may result from the extraction of substantially more components than the minimum needed for a stable decomposition (Abou-Elseoud et al., 2010; Kiviniemi et

After fMRI data preparation and preprocessing, a group statistical analysis is typically required to summarize RSN functional connectivity in one or more populations of interest and to search for possible regional differences between populations within selected RSNs. In many cases, population-level studies based on ICA use a two-level approach, first running single-subject ICA and then combining the components into a second-level group (random effects) analysis; in order to match components between subjects clustering and spatial correlation techniques are used (Esposito et al., 2005; Schopf et al., 2011; Wang & Peterson, 2008). This strategy provides maximal power to model subject-level structured noise (Cole et al., 2010) and has the important advantage of capturing unique spatial and temporal features of the subjects' data set even if the signal to noise ratio (SNR) is substantially lower in some subjects compared to other subjects. The disadvantage of this approach is that the components that are matched across subjects are not necessarily

As an alternative to clustering, temporal (Calhoun et al., 2001a; Varoquaux et al., 2010) and spatial (Svensen et al., 2002) concatenation as well as "tensorial" (Beckmann & Smith, 2005; Guo & Pagnoni, 2008) data aggregation schemes have been previously examined to perform only one ICA decomposition, thereby circumventing the problem of a "first-level" component matching. The most used aggregate group ICA approaches (Calhoun et al.,

extracted in the same way for each subject of a group (Erhardt et al., 2010).

connections that directly (monosynaptically) or indirectly (multisynaptically) make the ongoing communication physically possible (Greicius et al., 2009; van den Heuvel et al., 2009). Thereby, it is equally important to clarify whether the observed RSN functional connectivity is mediated by direct or indirect structural connections, e. g. by combining rsfMRI with diffusion tensor imaging (DTI), an MRI technique that allows the study of white matter fiber bundles.

By far, the most studied RSN in the clinical and research neuroimaging community is the DMN (Greicius et al., 2003; Raichle et al., 2001). This network has attracted considerable interest in the neuroscience community for its possible role as the baseline cognitive state of a subject and its link to memory and executive functions in normal and pathological conditions. In fact, the DMN normally includes the ACC and PCC regions, known to be involved in attention-related processes (Badgaiyan & Posner, 1998) and often detectable as transiently or consistently deactivated during many different types of cognitive tasks (McKiernan et al., 2003). For this reason, Raichle et al. (2001), who first targeted this type of brain activity with positron emission tomography (PET) imaging, have introduced the concept of "default-mode" activity and attempted to differentiate a "cognitive" baseline state from a "general" resting state in human brain. Thereafter (Greicius et al., 2003) the DMN has been often conceptualized as a "stand-alone" function or system to be analyzed with data models specifically oriented to functional connectivity (Bullmore et al., 1996). Within the DMN, the PCC node, one of the most intensively interconnected regions in the whole brain (Cavanna & Trimble, 2006; Hagmann et al., 2008), seems to mediate all the intrinsic functional connectivity of the brain (Fransson & Marrelec, 2008). Indeed, the PCC plays an essential role in all types of introspective mental activity, ranging from immediate suppressing of distracting thoughts to avoid mistakes (Li et al., 2007; Weissman et al., 2006) up to modulating rethinking about the past to imagine the future and awareness (Buckner et al., 2008).

The VIS network involves regions in the striate, peri-striate and extra-striate visual cortex, which are normally activated by a visual task. This network extends from the lingual and fusiform gyri (i. e. V1 to V4) up to the occipito- and middle temporal regions (i. e. MT/V5), even if, in some reports, regions belonging to the primary and secondary visual system are shown to belong to separate visual RSNs (see, e. g., (Rocca et al., 2011)).

The fronto-parietal (or "executive-attention") network (FPN), sometimes found to be lateralized (i.e., right and left FPN), is also relevant for cognition. Particularly, the FPNs seem to be central for cognitive processing as they involve regions such as the dorsal frontal and parietal cortices potentially overlapping with the dorsal attention network which is known to mediate executive control processing (Corbetta & Shulman, 2002).

The SMN includes regions in the precentral and postcentral gyrus and in the supplementary motor area, all regions that both anatomically and functionally correspond well to motor and sensory areas, e. g. activated during a finger tapping task (Biswal et al., 1995).

The AUD network involves regions in the auditory cortex, which are normally activated by an auditory task. This network extends from the Heschl's gyrus to the superior temporal gyrus and the insula and has been also reported as one or two RSNs (see, e. g., (Damoiseaux et al., 2006)).
