**3. Methodology for the analysis of the resting-state networks**

RSN can be observed using several functional connectivity analysis tools. A straightforward approach entails extracting the extracting the time-course of the BOLD signal from a pre-

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

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

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

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

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

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

RSN can be observed using several functional connectivity analysis tools. A straightforward approach entails extracting the extracting the time-course of the BOLD signal from a pre-

shown to belong to separate visual RSNs (see, e. g., (Rocca et al., 2011)).

known to mediate executive control processing (Corbetta & Shulman, 2002).

**3. Methodology for the analysis of the resting-state networks** 

and sensory areas, e. g. activated during a finger tapping task (Biswal et al., 1995).

matter fiber bundles.

al., 2008).

et al., 2006)).

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 data sets, thereby avoiding the possibility of bias.

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

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 al., 2009).

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

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.,

Neuronal Networks Observed with Resting State

Wong et al., 1995; Wils et al., 2010).

Functional Magnetic Resonance Imaging in Clinical Populations 115

are strongly reduced can be easily linked to most existing animal models of ALS explaining motor neuron degeneration both at the cellular and molecular levels (Dal Canto et al., 1995;

Fig. 2. ALS disease effects in the SMN. Upper panel: F-map of statistically significant disease effects within the SMN network (P=0.05, cluster-level corrected) overlaid on the average Talairach-transformed T1 image (coronal and axial cuts). Lower panel: Scatter plots of the regional ICA z-scores vs age in the R-PMC (left) and in the L-PMC (right). PMC = primary

The RFPN network is also partially suppressed in ALS patients. Figure 3 shows the localization of two regions within this network, in the superior frontal gyrus (SFG) and in the supra-marginal gyrus (SMG), where the network-specific RS-fMRI fluctuations resulted suppressed in ALS compared to controls. These effects in a cognitive executive network like the RFPN likely reflect a rather typical frontal cortex dysfunction observed in ALS patients (Abrahams et al., 1996; Hatazawa et al., 1988; Rule et al., 2010; Vercelletto et al., 1999). Observing RSNs in ALS patients over an extended range of age has highlighted the possible interaction between aging and neurodegeneration (Tedeschi et al., 2010). Previous work has reported a significant effect of aging on DMN regions in the normal population (Esposito et al., 2008; Grady et al., 2006; Greicius et al., 2004; Koch et al., 2009; Persson et al., 2007). In ALS patients, the DMN network has shown an age-by-disease interaction effect in the PCC (figure 4), with the strength of the RS-fMRI fluctuations relatively increased rather than reduced with increasing age (and disease duration). In addition, there was also a group-byage interaction effect in RFPN, and more precisely the middle frontal gyrus (MFG) (figure

motor cortex. ALS = amyotrophic lateral sclerosis patients. CTL = control subjects.

2001a; Zuo et al., 2010) are based on temporal concatenation and assume "common" ICA maps for all subjects in the first level analysis. A population analysis is then performed retrospectively determining the individual ICA components from the group ICA components. Thereby, all these methods implicitly assume that a given component is really present with exactly the same layout in all the subjects.
