Development of large-scale functional networks from birth to adulthood: a guide to neuroimaging literature

The human brain is organized into multiple distributed functional brain networks that can be measured at multiple spatiotemporal scales

Coordinated neuronal activity between anatomically disparate regions is an essential feature of human brain function. Across all stages of postnatal development, brain activity to a large degree is consolidated within so-called “resting-state networks” (RSNs). RSNs are defined as distinct modules of regions that exhibit highly synchronized activity even in the absence of external stimuli. RSNs have proven to be highly reproducible in the adult brain (Doucet et al., 2011; Gordon et al., 2016; Power et al., 2011; Yeo et al., 2011) and have become an influential framework for interpreting functional and structural neuroimaging data. Patterns of functional connectivity (FC) within and between the major RSNs are increasingly understood as intrinsic properties of brain function, given that they strongly predict patterns of interregional co-activation across different tasks (Cole et al., 2014; Smith et al., 2009) and are associated with task-relevant behavioral performance (Cole et al., 2012; Lewis et al., 2009).

Interregional FC arises from anatomical projection strengths (Messe et al., 2014; Shen et al., 2012), correlated gene expression (Richiardi et al., 2015), and synaptic receptor densities (Turk et al., 2016; van den Heuvel et al., 2016b)–properties that, in turn, undergo experience-dependent and activity-dependent modulation over the lifespan (Huttenlocher, 2002; Markham and Greenough, 2004; Scholz et al., 2009). This intricate structural-functional interplay underscores that the ontogeny of functional networks likely reflects programmed neurodevelopmental events (e.g. neurogenesis, cell death, myelination, pruning, synaptic plasticity, and glial development) (Innocenti and Price, 2005; Stiles and Jernigan, 2010) and that FC aberrations may point towards the etiological bases of neuropsychiatric disorders (e.g. (Swartz and Monk, 2014; van den Heuvel et al., 2016a)). Thus, exploring the normal trajectory of functional network development and its relationships with these underlying biological processes should be a central mission of both basic and clinically relevant neurodevelopmental research.

Current views regarding network development have been heavily informed by resting-state functional connectivity MRI (rs-fcMRI), which measures the correlations in spontaneous, low-frequency activity between investigator-defined regions. This review will accordingly provide in-depth discussions of the developmental rs-fcMRI literature, its limitations, and unique insights gained from complementary approaches, especially electroencephalography and magnetoenceophalography (EEG/MEG). We will begin by discussing emerging themes in the analysis of functional network data. We will then synthesize current findings regarding FC changes from birth to adulthood, and place them within the context of co-occuring macro- and micro-scale structural modifications. Finally, we will discuss some of the implications with regard to neurodevelopmental disorders and highlight several crucial knowledge gaps that are likely to guide research efforts in the near future.

Analytic approaches to functional connectivity are evolving rapidly

Correlated resting-state activity was first identified in bilateral sensorimotor cortex by (Biswal et al., 1995), followed soon after by similar findings of bilaterally correlated activity in early visual and early auditory cortex (Cordes et al., 2001; Kiviniemi et al., 2000; Lowe et al., 1998). These discoveries were made using “seed” regions outlined based on a priori anatomical boundaries. Correlations were then quantified between the seed region’s fMRI timecourse and all other voxels of the brain, producing an FC map. More data-driven approaches were developed based on the concept of independent components analysis (Beckmann and Smith, 2004), which decomposes brain activity into a set of spatial maps (i.e. components) that minimally overlap. ICA maps reproducibly reveal early visual, early auditory, and primary sensory/motor components, as well as other components comprising distributed portions of the frontal, parietal, and temporal association cortices (Damoiseaux et al., 2006).

Building upon these advances, the neuroimaging community has increasingly focused on interrogating the brain as an integrated complex system. In this context, graph theory has emerged as a powerful new approach. A ‘graph’ is simply a network of things, referred to as nodes, and the connections between those things, referred to as edges. The network’s behavior can thus be modeled as a set of properties that emerge from the network’s unique global structure and local features (Sporns, 2014). Graph theoretic studies typically construct functional brain networks using regions as nodes and correlations in activity as edges. Past reviews have already written eloquently about the rationales for applying this approach to study development (Power et al., 2010; Vertes and Bullmore, 2015), which we will only briefly summarize and expand upon here.

First, most individual cognitive abilities arise not solely from a particular brain area, but from networks of activity spanning multiple distributed regions (Petersen and Sporns, 2015). From social cognition in infancy (Eggebrecht et al., 2017) to cognitive control and decision-making in adolescence (Dwyer et al., 2014; van Duijvenvoorde et al., 2015), specific cognitive capacities co-evolve with complex network effects within and between different RSNs, throughout postnatal development Second, structural connectivity shapes and constrains functional networks across the lifespan (Betzel et al., 2014; Hagmann et al., 2012; Vertes and Bullmore, 2015). Cross-modal network analysis is therefore crucial for elucidating normative mechanisms of cognitive development and the pathophysiology of neurodevelopmental disturbances. As a proof-of-principle, recent studies have found that even purely local, experimentally induced perturbations (e.g. exogenous stimulation or inactivation) result in widespread, complex neurophysiological changes, which are at least partly explainable as local interactions with global network structure (Andoh et al., 2015; Grayson et al., 2016; Gu et al., 2015a; Misic et al., 2015). By extension, one could reason that developmental modifications in brain FC, whether normative or pathological, are also best understood via an emerging network science that merges structural and functional connectivity data. Third, graph-based methods are flexible and generalizable. Networks can be measured at various temporal scales depending on imaging methodology, at various conditions (during rest or task completion), and can be compared with corresponding networks of structural connectivity or gene co-expression. A graph theoretic approach is therefore well-suited for multimodal investigations linking brain function across temporal scales or examining structural drivers of functional development.

Community Structure

A defining feature of human brain activity is “modularity,” which refers to an unexpectedly high level of within-RSN FC relative to between-RSN FC (Lohse et al., 2014). “Community detection” algorithms partition the brain into distinct communities, or modules, by maximizing this modularity quotient. The adult brain has several modules at the coarsest level that are remarkably reproducible despite slight methodological differences across studies (Doucet et al., 2011; Gordon et al., 2016; Power et al., 2011; Yeo et al., 2011). These modules include 1) segregated, non-distributed modules for sensory/motor activity, such as in early visual cortex and in somatomotor cortex, 2) the default-mode network (DMN), comprising the medial prefrontal, posterior cingulate/precuneus, angular gyrus, and medial temporal cortex; the DMN is considered “task-negative” as it has been linked to internal mentation and is known to deactivate during processing of external stimuli (Buckner et al., 2008), and 3) three large, distributed modules for task-positive cognition: a) dorsal attention, involved in eye movements and attentional orienting, b) the cingulo-opercular salience module, involved primarily in attentional maintenance, and c) the frontoparietal executive module, important for task-switching. At finer resolutions, several other modules can be identified (Doucet et al., 2011; Gordon et al., 2016; Power et al., 2011; Yeo et al., 2011), including hierarchical visual subdivisions, topographical somatomotor subdivisions, an auditory module, a limbic module, a ventral attention module, and other smaller subdivisions of the executive, salience, and default-mode modules. See Figure 1A & B for a graphical summary.

The brain’s modular composition is believed to help segregate information processing between distinct sensory modalities or cognitive architectures. It remains a fundamental challenge to understand what other network properties help integrate disparate data streams in order to meet complex task demands (Cohen and D’Esposito, 2016; Petersen and Sporns, 2015). As discussed throughout this review, the themes of segregation and integration are core concepts in understanding the brain’s maturational processes. Thus, further analytic exploration is needed to isolate the critical network components that facilitate segregation, integration, or efficiency. Below we briefly describe some local network measures toward this end.

Node Measures

The most straightforward nodal description is its strength (Rubinov and Sporns, 2010), defined as the sum of a node’s connection weights. Other centrality measures quantify higher order measures of influence. For instance, betweeness centrality (Rubinov and Sporns, 2010) computes the frequency with which a node lies on the shortest paths that connect any other nodes, a property that theoretically enables efficient global communication. By far, strength and betweenness have been the most popular hub metrics within neuroimaging, although others such as eigenvector centrality (Lohmann et al., 2010; Zuo et al., 2012) and node communicability (de Reus and van den Heuvel, 2014; Mantzaris et al., 2013) have seen useful application as well. These metrics, however, require careful consideration of what the network’s edge weights signify (see (de Reus and van den Heuvel, 2014; Grayson et al., 2016; Misic et al., 2015) for discussions about the neurobiological relevance of shortest paths). Investigators must especially consider whether connections are statistical or structural in nature. For example, a node with structural links to many different RSNs (and thus, high structural network centrality) may have weak statistical associations overall (and thus, low functional network centrality). It has therefore been proposed to consider functional hubs as nodes with integrative capacity across distinct modules (Power et al., 2013). This goal is accomplished in part by the participation coefficient (Guimera and Nunes Amaral, 2005), which computes the between-module connectivity and therefore is particularly useful for functional networks (see Figure 1C).

Another way of identifying nodal integration is to compare functional against structural networks (see Figure 1E). Structural networks in the mammalian brain exhibit an important phenomenon known as “rich-club” organization, whereby hubs (i.e. nodes with high strength) link disproportionately to other hubs. Rich-club nodes tend to span multiple RSNs, thereby facilitating integration between them (Sporns, 2013; van den Heuvel and Sporns, 2013). Similarly, structural hubness has been linked to integration of global brain activity (Grayson et al., 2016).

Finally, node clustering coeffiicients and local efficiency (Rubinov and Sporns, 2010) are popular ways of measuring node specialization – i.e., integration specifically within a node’s local environment. The clustering coefficient quantifies the combined strength of connectivity between a given node’s immediately connected neighbors, whereas the closely related local efficiency computes the average (inverse) path length between a node’s neighbors.

Global Network Properties

A network’s capacities for globally efficient communication and integration can be defined formally as properties. The global efficiency is measured by averaging the (inverse) shortest path length between all node pairs. Greater global efficiency in structural and functional networks has been linked to higher IQ, attentional capacity, and working memory in healthy young adults (Cohen and D’Esposito, 2016; Kitzbichler et al., 2011; Li et al., 2009; Stanley et al., 2015; van den Heuvel et al., 2009), supporting the notion that it confers advantages for complex cognitive processing. When networks exhibit both high global efficiency and high average clustering coefficients, they exhibit “small-world” organization (Watts and Strogatz, 1998). This property, which is ubiquitous across vastly different scales of investigation in the brain, is thought to emerge from a biological need to balance the benefits of integration against the costs of long-range wiring (Collin et al., 2013; Vertes and Bullmore, 2015). Furthermore, these distinct network effects, i.e. segregated versus integrated neural processing, are interwoven features of cognition. Both effects in tandem are crucially involved in cognitive control (Dwyer et al., 2014) and executive functioning (Reineberg et al., 2015), and are differentially associated with performance on simple motor tasks versus complex working memory tasks (Cohen and D’Esposito, 2016). This review will highlight extensively these competing network influences in the context of development. Finally, network integration can also be defined alternatively. Similar to the participation coefficient of a given node, one can compute the network’s global integrative capacity by taking into account the network modules and the interdependencies between them (de Reus and van den Heuvel, 2014; Tononi et al., 1994).

Overall structural and functional organization

The human brain expands at an explosive rate during the first few years of life (see Figure 2A), doubling in volume during the first year and reaching 80–90% of adult volume by age 3 (Courchesne et al., 2000; Knickmeyer et al., 2008). This same time period is marked by rapid and widespread cortical synaptogenesis, followed by a protracted period of synapse elimination and cell loss which carries into adulthood (Huttenlocher, 2002; Innocenti and Price, 2005). The brain’s major fiber pathways become consolidated through myelination, though their presence is largely established prenatally (Stiles and Jernigan, 2010). It is also worth noting that both progressive and regressive changes in structure occur simultaneously during infancy (Gilmore et al., 2012; Scott et al., 2015), despite overall growth in volume and connectivity. These first few years of life, therefore, coincide with the formation of the brain’s foundational anatomical circuitry, paralleling the genesis of an intricate functional architecture.

RSNs are already present during infancy, and potentially to some extent in utero (Thomason et al., 2013). Seed-based and component-based approaches demonstrate the existence of robust, bilateral segregated networks for somatomotor, primary auditory, primary visual, and extrastriate visual cortex (Fransson et al., 2007; Gao et al., 2015a; Kiviniemi et al., 2000; Lin et al., 2008; Liu et al., 2008; Smyser et al., 2010). These sensory networks undergo subtle refinements and strengthening over the first two years of life, and by age two bear substantial resemblance to their adult counterparts (Gao et al., 2015a; Lin et al., 2008). Distinct components throughout heteromodal association cortex have also been described in infants, and are believed to represent prototypical elements of distributed, higher-order cognitive networks identifiable in adults (Graham et al., 2015). For instance, in adults the DMN is a tightly integrated circuit comprising several distinct and distant areas of cortex (as described in Community Structure). In neonates and very young infants, these individual compartments demonstrate locally correlated activity, but fail to synchronize into a coherent network (Damaraju et al., 2014; Fransson et al., 2007; Gao et al., 2009; Smyser et al., 2010; Wylie et al., 2014). Over the course of the first two years of life, these regions undergo a precise evolution to gradually bring the network ‘online’ (Damaraju et al., 2014; Gao et al., 2013; Gao et al., 2009). By two years of age, the network bears adult-like qualities, although coupling between distant anterior and posterior nodes of the network remain relatively low. Similar spatial and temporal properties characterize the emergence of distributed connectivity in the dorsal attention (Gao et al., 2013), salience (Alcauter et al., 2015a; Gao et al., 2015a; Gao et al., 2015b) and frontoparietal executive (Gao et al., 2015a; Gao et al., 2015b) networks. Collectively, these findings suggest that sensory networks become established at a much earlier age relative to those implicated in higher-level cognition (Gao et al., 2016). These trends are consistent with the presence of basic somatosensory and visual functions at birth, and help contextualize findings that primary sensory regions are generally the first cortical areas to reach plateaus or peaks in growth (Geng et al., 2016; Gilmore et al., 2012; Lyall et al., 2015; Scott et al., 2015).

Development of community structure from birth

Despite these insights, network topology in early development remains poorly understood. What is the collective community structure of the brain at birth? How does it evolve during early and late infancy, and how is it distinguished from the now well-described adult community structure?

The current nascent literature on infant network topology (see Figure 2B) suggests that a robust community structure does exist, albeit in a primitive form. In neonates, two independent rs-fcMRI studies found evidence for modules comprised largely of anatomically proximate areas (De Asis-Cruz et al., 2015; van den Heuvel et al., 2015b). One recent report compared community structure in adults against that in 1–2 year olds using closely matched methods (Eggebrecht et al., 2017), providing the first direct evidence for numerous anatomically localized modules in the infant brain and their relationship to adult functional organization (Figure 2C). Many of these modules clearly reflect anterior/posterior/temporal subdivisions of the adult salience, frontoparietal, DMN, dorsal attention, and ventral attention modules. Another distinction seen is that infant modules tend to group together anatomically proximate ROIs that have divergent roles in adulthood. Community structure assessed at two years of age (Alcauter et al., 2015b; Gao et al., 2015b) indicates the presence of distinct modules encompassing early auditory, early somatomotor, and early visual areas, and potentially, other prototypical modules similar to the canonical salience, frontoparietal, and DMN seen in adults. These latter modules also exhibit substantial growth of distributed synchronization and homotopic coupling from birth through the first two years of life (Gao et al., 2015b; Homae et al., 2010), concomitant with decreasing between-module connectivity (i.e. greater functional specialization). Taken together, these studies point towards several trends that are consistent with results from seed and component based approaches – 1) modular patterns at birth take a coarse, primitive form that is heavily influenced by local anatomy, 2) this structure undergoes rapid postnatal refinement, leading to progressively stronger coupling between distributed links that facilitate modular specialization, and 3) these enhancements result in the establishment of stable, primary sensory networks well before networks involved in higher-level cognition appear substantially adult-like.

Regional hub metrics

Hub configurations are robustly evident after birth (De Asis-Cruz et al., 2015; Fransson et al., 2011; Gao et al., 2011), but undergo substantial changes postnatally. Multiple reports have found that high strength and betweenness hubs in adulthood are largely located within within the default mode, attentional, sensorimotor, and visual areas (Cao et al., 2014; Fransson et al., 2011; Grayson et al., 2014; Hwang et al., 2013), whereas neonatal hubs primarily involve early sensorimotor, visual, and limbic regions (De Asis-Cruz et al., 2015; Fransson et al., 2011; Gao et al., 2011). The first two years of life reflect a gradual decrease in the hubness of sensorimotor areas and increased involvement of various association cortical areas, especially the medial frontal and parietal portions of the default mode (Gao et al., 2011). This pattern of hub strengthening bears substantial spatial and temporal overlap with the pattern of increasing within-module synchronization. This is consistent with the notion that node strength and betweenness in functional networks may serve, at least to some extent, as proxies for modular size and synchronization (Power et al., 2013). Thus, there continues to be much less known regarding what regions become more involved in the integration of different information processing streams (e.g. sensory, cognitive, and emotional) over time. Such anatomical areas are likely critical for the development of complex cognitive and social abilities, and are potentially uniquely susceptible to early environmental influences (Graham et al., 2015). It is also clear that several prenatal factors (e.g. immune and endocrine systems) strongly affect postnatal network trajectories (Canetta et al., 2016; Scheinost et al., 2016b). Future work examining node participation coefficients in functional networks and other centrality measures in structural networks (see network measures) promises to shed greater light on the emergence of network integrators and their potential susceptibility to pre- and post-natal factors.

Global network metrics

Other important topological properties are also robustly evident at birth. Small-worldness (De Asis-Cruz et al., 2015; Fransson et al., 2011; Gao et al., 2011), as well as its constituent properties of global efficiency and average clustering coefficient, are already high in neonates. These properties only appear to marginally increase during the first year of life, plateauing thereafter until at least the end of the second year (Gao et al., 2011). Concomitantly, long-distance links are preferentially strengthened during the first year of life, leveling off thereafter (Gao et al., 2011). Rich clubs are also evident in both structural and functional networks in neonates along with both high functional modularity and integration capacity between modules (Ball et al., 2014; Scheinost et al., 2016a; van den Heuvel et al., 2015b). These findings suggest that the development of a small-world architecture, as well as the capacity for globally efficient communication and integration, largely occurs during the first year of life and prenatally, a finding that is supported by multimodal structural and functional network analysis in preterm infants (Ball et al., 2014; van den Heuvel et al., 2015b). This (very) early developmental trajectory stands in contrast to the more protracted progression towards increasingly segregated and refined modular architecture which continues into early childhood and beyond.

Considerations in infant rs-fcMRI

A crucial caveat to the above observations is the potential for unidentified sources of artifact in infant studies (also see Methodological challenges and recommendations). In the case of identified networks in infants it is important to note that different community detection algorithms can provide unique results and all are not appropriate for any given experimental condition (Gates et al., 2016). Other considerations include early postnatal changes in cerebral vasculature, and altered state of consciousness when scanning very young participants. fMRI relies on hemodynamic measurements (usually BOLD), and there is good reason to suspect that infancy is accompanied by specific changes in neurovascular coupling (Hagmann et al., 2012). Human and animal studies indeed indicate differences in stimulus-evoked neurovascular coupling between infants and older individuals, but whether these differences impact rs-fcMRI measurements has yet to be determined (Graham et al., 2015). Infant imaging studies are also predominantly done during natural sleep (e.g. (Liu et al., 2008)) or sedation (e.g. (Fransson et al., 2007)). A natural benefit is the attenuation of micromovements and their attendant signal artifacts. However, altered state of arousal and consciousness might pose important confounds when comparing network findings to those seen in conscious children or adults. This concern is partly mitigated by findings of similar global functional network topology in conscious versus unconscious state in adults (Spoormaker et al., 2010), but key differences have been found in corticocortical (Heine et al., 2012; Horovitz et al., 2009; Spoormaker et al., 2010) and thalamocortical (Spoormaker et al., 2010) connectivity. Different sleep stages also correspond with different neurophysiological phenomena, which impact rs-fcMRI measurements (Laumann et al., 2016; Spoormaker et al., 2010). Until additional work can more definitively tease apart artifactual vs. neural contributions to the fMRI signal or awake versus sleep affects in these early postnatal years, developmental changes in modular partitions and other network changes must be viewed with at least some caution.

General observations

Many of the network changes seen during infancy reflect long-term trajectories that extend into childhood and adolescence. As with the infant literature, much of the literature in children and adolescents has focused on the regions that define the adult DMN. Seed and component approaches have consistently found that connectivity between these areas (and indeed, connectivity within other cognitive RSNs) continues to strengthen from early childhood throughout development, especially with respect to long-range anterior-posterior links (de Bie et al., 2012; Fair et al., 2008; Sato et al., 2014; Sherman et al., 2014; Supekar et al., 2010).

More generally, one of the more reproducible findings of child and adolescent development (until very recently) had been an apparent shift from a local to global organization. Prior to 2012, multiple studies had found that functional connectivity between anatomically proximate ROIs gradually decreased, whereas long-range functional connections, especially anterior-posterior links, gradually increased until adulthood (Dosenbach et al., 2010; Fair et al., 2009; Kelly et al., 2009; Supekar et al., 2009). In addition, it appeared that the modular assignments of regions shifted. In childhood, regional assignments appeared to strongly depend on local anatomy. With age, regions gradually became more defined by their respective functional roles (Power et al., 2011; Power et al., 2010). These trends pointed to coherent changes that involved a dampening of non-specific local spread of activity and strengthening of long-range links, facilitating cognitive development through modular specialization and selective cross-network integration (Bunge and Wright, 2007; Fair et al., 2009; Uddin, 2011; Uddin et al., 2011). Distance-dependent effects were even strong enough to enable reasonably accurate predictions of an individual subject’s age on the basis of a single resting-state scan (Dosenbach et al., 2010).

However, recent concerns about the influence of head-motion artifact have raised some doubts about the interpretability of such broad trends. In particular, it has been revealed that motion increases non-specific local coupling and decreases long-range coupling in ways that are similar to differences found between children and adults (Fair et al., 2012b; Power et al., 2012; Satterthwaite et al., 2012; Yan et al., 2013). When processing techniques intended to remove motion artifact are applied, developmental effects of this type – both in terms of overall local versus distributed coupling and in terms of changing modular assignments – are substantially attenuated, though still present to a limited extent (Fair et al., 2012b; Power et al., 2012; Satterthwaite et al., 2012). As motion-denoising strategies continue to evolve, such findings will need continual re-evaluation. It is likely, however, that functional network maturation follows more precise spatiotemporal trajectories than previously understood. Indeed, there are now at least 3 reports that suggest the network structure in children as defined by community detection (see Figure 3) are largely similar to the adult (Fair et al., 2012b; Marek et al., 2015; Power et al., 2012). Despite this “adult like” level of organization, refinements within and between systems does appear to occur over time, although details to this end are still being worked out (Fair et al., 2012b; Gu et al., 2015b; Marek et al., 2015).

Much of this new work is finding that systems with different functional roles are characterized by substantially different trajectories (Gu et al., 2015b; Marek et al., 2015). For example, (Gu et al., 2015b) found that sensorimotor systems, already well-segregated in childhood, displayed little change into adulthood. On the other hand, networks involved in task-positive cognitive function (e.g. salience, frontoparietal executive, and attentional networks), being generally well-connected with other systems, became increasingly segregated and exhibited increasingly variable spatial structure. The DMN was found to exhibit both increasingly synchronized within-network and between-network connections. The authors contend that their results are consistent with notions that network development facilitates the ability to traverse through more diverse cognitive states in order to accomplish increasingly complex task demands (Bassett et al., 2011; Braun et al., 2015). Similar, albeit not identical developmental findings, are observed by (Marek et al., 2015)), who found increased cross-network integration between the cingulo-opercular/salience and somatomotor modules over the age of 10–26, which they showed to be associated with the development of cognitive control. In agreement with these data, two additional reports also had findings that suggested that functional restructuring over this timespan disproportionately involves the cingulo-opercular and somotomotor modules (Fair et al., 2012b; Grayson et al., 2014), and potentially specific links between them (Grayson et al., 2014) (see Figure 4). Therefore, there is recent convergence of data suggesting that network structure continues to evolve after late childhood in ways that are not primarily distance-dependent; rather, complex but specific changes may coincide with the development of specific cognitive abilities. Additional follow-up work using up-to-date methods for artifact prevention and removal, and more sensitive longitudinal designs, will be crucial for expanding on these important new findings.

EEG/MEG findings

Resting-state networks obtained using electrophysiological recording (EEG or MEG) provide greater temporal resolution (<100Hz) relative to rs-fMRI (<0.1Hz) and therefore offer complementary insights about functional development. One of the classical and well-reproduced findings is that spectral power (i.e. amplitude of activity within different temporal frequency bands) decreases continually from 5–30 years of age across most frequency bands, but particularly in the slow-wave (0.5–7Hz) (Miskovic et al., 2015; Rodriguez-Martinez et al., 2015; Smit et al., 2012; Whitford et al., 2007). There is also a progressive shift in peak amplitude towards higher frequency activity (Miskovic et al., 2015; Rodriguez-Martinez et al., 2015; Smit et al., 2012; Whitford et al., 2007). Interestingly, frontal and parietal cortical gray matter (but not white matter) volume decreases across this same age bracket, and exhibits curvilinear trajectories that are closely correlated to the loss of slow-wave power in these same areas (Buchmann et al., 2011; Smit et al., 2012; Whitford et al., 2007). These coincident trajectories have provided key support for the hypothesis that developmental changes in baseline electrophysiological signatures may primarily be subserved by reduction of neuropil and progressive synapse elimination (Buchmann et al., 2011; Smit et al., 2012; Whitford et al., 2007).

EEG/MEG functional connectivity networks can also be constructed from synchronized activity within different frequency bands. Reports of resting-state EEG/MEG networks in infancy are scant (although see later section on temporal dynamics), but a number of network studies have been conducted regarding later development. It is important to note that these studies are generally unable to map activity to precise brain areas due to challenges in source localization, and often treat the scalp sensors as the network nodes. Nonetheless, there are several noteworthy findings. From roughly 5 years of age until adulthood, theta (~4Hz), alpha (~10Hz), and beta (~20Hz) oscillation networks exhibit concurrent increases in clustering coefficient (Boersma et al., 2011; Smit et al., 2012) and strength of long-range anterior-posterior and interhemispheric connections (Barry et al., 2004; Miskovic et al., 2015; Smit et al., 2012; Srinivasan, 1999). There is also consensus that network structure becomes increasingly non-random across this timespan (Boersma et al., 2011; Miskovic et al., 2015; Smit et al., 2012). Somewhat unexpectedly, two reports found path length to also increase with age (Boersma et al., 2011; Smit et al., 2012), although more recent work demonstrates age-related reductions (Miskovic et al., 2015). While this work is in its infancy, these studies thus far suggest some consensus between electrophysiological and MRI-based findings, implying that changes in brain organization are at least partially preserved across temporal scales of investigation.

Regional hub metrics

Returning to the MRI literature, there is convergent evidence that the locations of high-strength and high-betweenness hub regions appear to be stable after 5 years of age (Cao et al., 2014; Grayson et al., 2014; Hwang et al., 2013; Wang et al., 2012; Wu et al., 2013; Zuo et al., 2012). However, there is also evidence for regional refinement of hub properties. Node strength and centrality of subcortical regions have both been reported to decrease in late childhood and adolescence (Sato et al., 2015; Supekar et al., 2009). Among cortical areas, some have reported increasing hubness of frontal nodes (Cao et al., 2014; Wu et al., 2013) while in other lobes there have been mixed reports of increasing (Cao et al., 2014; Grayson et al., 2014) and decreasing (Wu et al., 2013; Zuo et al., 2012), or stable (Hwang et al., 2013) hubness. At first glance, these mixed or non-significant findings might suggest that regional properties of cortical hubs are largely established during very early development. However, given that different networks display varying trajectories for within-network and between-network connectivity (Gu et al., 2015b; Marek et al., 2015), renewed examination of nodal hub-like properties such as the participation coefficient (which quantifies within vs. between module strength) may be warranted (e.g. see (Marek et al., 2015)). More detailed network analyses have revealed that specific connections to cortical hub regions increase during childhood and adolescence (Grayson et al., 2014; Hwang et al., 2013) resulting in increased rich club coefficients in adulthood (Cao et al., 2014; Grayson et al., 2014). Regional properties have also been found to exhibit nonlinear relationships with age (Wang et al., 2012), emphasizing the importance of mapping trajectories with more detailed temporal precision.

FMRI signal artifacts

Rs-fcMRI investigators must be mindful that the fMRI signal is susceptible to a number of spurious sources of variance due to hardware instabilities, respiratory and cardiac signals, and head motion artifacts (Jo et al., 2010; Power et al., 2016). Effects of respiratory and cardiac activity can be attenuated through the use of temporal filtering, ICA-denoising, or directly acquiring these signals and removing them post-hoc. There is ongoing debate regarding the adequacy of these techniques (Jo et al., 2010; Power et al., 2016), and more work may be warranted to determine optimal ways to account for age-related physiological confounds. In addition, head motion has become a substantial concern since 2012 when a trio of papers pointed to systematic motion-related artifacts in rs-fcMRI (Power et al., 2012; Satterthwaite et al., 2012; Van Dijk et al., 2012). Chief among these findings was distance-dependent effects; head motion increases functional correlations between spatially proximal nodes and decreases correlations between distant nodes (Power et al., 2012). Motion tends to be greater in patient populations and younger children, posing a fundamentally important challenge for developmental research. Considerable work and debate is now ongoing to determine the optimal ‘denoising’ approaches during data processing (Patriat et al., 2016; Power et al., 2014). The most contentious debate probably concerns the use of global signal regression (GSR) – arguably the most effective motion-denoising technique currently available (Burgess et al., 2016; Power et al., 2015; Yan et al., 2013). Recent data reinforces concerns about artifactual contributions to the global signal in conventional 3T rs-fcMRI acquisitions sensitive to BOLD contrast (Grayson et al., 2016; Power et al., 2016), but also partly substantiates concerns regarding GSR under less conventional, high signal-to-noise scenarios where motion can be physically restricted (Grayson et al., 2016). Thus, GSR is likely to be a useful tool in developmental rs-fcMRI studies, where good control of such artifact is exceptionally important, but this debate may need to be revisted as methodologies in acquisition continue to advance. Moving forward, there is strong consensus about the importance of preventing head motion during acquisition and rigorously quantifying motion effects in study data (Goto et al., 2016; Power et al., 2015; Siegel et al., 2016).

We also wish to emphasize that future work using strict artifact removal are an increasing need for neuroimaging studies quite broadly – well beyond functional connectivity studies. Strong artifacts due to motion, in addition to their relevance for fMRI, have been noted in diffusion weighted imaging (Roalf et al., 2016; Yendiki et al., 2014) and cortical thickness measurements (Reuter et al., 2015; Savalia et al., 2017). Motion artifacts appear to be tightly related to clinical factors (Fair et al., 2012b) and a whole host of behavioral phenotypes and metrics (Siegel et al., 2016). Some of the techniques currently being used to remove these artifacts are non-optimal (Burgess et al., 2016; Goto et al., 2016). Thus, close examination of many of the developmental MR imaging findings is still warranted.

Parcellation choice

Developmental network researchers also must make important decisions when defining the nodes (i.e. functional areas) under investigation. There is currently no consensus regarding an optimal set of criteria for defining an areal parcellation (Craddock et al., 2013; Glasser et al., 2016a; Gordon et al., 2016), and even less known about the stability of these parcellations at younger ages.

Parcellation density (i.e. number of regions) is a crucial consideration. Global topological metrics for structural and functional networks show stability across different parcellation schema of similar density, but vary considerably as a function of network size (de Reus and van den Heuvel, 2013; Fornito et al., 2010; Zalesky et al., 2010). In addition, with functional connectivity data some parcellation schemes can obscure known community structure. For example, regions or parcels that are overly large may encompass multiple distinct functional areas in ways that make community/network organization non-recognizable (Power et al., 2011). Functional network analyses that strongly relate to known network organization may therefore find signal homogeneity within parcels to be a useful optimization criterion for choosing a parcellation (e.g. see (Gordon et al., 2016) and (Glasser et al., 2016a).

In our opinion, the primary guiding factors in choosing a parcellation are based in fundamental principles of neuroscience. The cerebral cortex is both anatomically and functionally organized at many physical scales - starting at the level of single neurons and extending up to functional systems (Churchland and Sejnowski, 1992). At one scale are discrete regions of the cortex, known as functional areas. Functional areas possess unique internal structure (e.g. architectonics and topography). They also contain distinct combinations of inputs and outputs, which in turn group related areas into specialized networks (Felleman and Van Essen, 1991). These areal boundaries appear very early in development (Kostovic et al., 1995) and each area is thought to make a distinct contribution to information processing (Cohen et al., 2008; Felleman and Van Essen, 1991; Fuster, 2002; O’Leary and Nakagawa, 2002; Sur and Rubenstein, 2005). Ideally, parcellations for functional connectivity studies would closely reflect these areal boundaries in order to more clearly link network effects to the brain’s underlying information processing architecture. In animal studies (rodents and non-human primates), this is feasible through the use of comprehensive, histologically and functionally defined whole-brain areal parcellations (Bezgin et al., 2012; Lein et al., 2007; Paxinos et al., 2000; Saleem and Logothetis, 2007; Swanson, 2015; Van Essen et al., 2012). In humans, there is much less clarity regarding potential areal boundaries (outside of elementary cytoarchitectonic divisions described, for example, by (Brodmann, 1909) and (Economo and Koskinas, 1925)) since a comparable amount of comprehensive data has not been readily available in a common reference space. Several groups, however, are finding ways to utilize today’s non-invasive imaging tools to bring us closer to realizing these goals (Behrens et al., 2003; Glasser et al., 2016a; Gordon et al., 2016; Wig et al., 2014a; Wig et al., 2014b; Zhang et al., 2010).

Atlas or template choice

A topic related to parcellation choice is how to spatially normalize individual subjects onto a common reference template. Cortical folding patterns change continually from birth into adulthood, and especially during infancy. Infancy is also a time during which the image contrasts in various MR acquisitions changes dramatically. These changes pose challenges during template-based data processing steps such as spatial normalization and image segmentation, and are particularly problematic in studies that use template brains derived from only one age group and therefore rely on assumptions regarding similarity in shape characteristics of the brain across participant groups. The use of high-contrast, age-specific and study-specific template brain images for image registration is a well-validated approach to help overcome some of these confounds (Avants et al., 2010; Fonov et al., 2011), and is increasingly adopted in both structural and functional studies of (especially early) development (Eggebrecht et al., 2017; Fonov et al., 2011; Gao et al., 2015a; Scott et al., 2015). Surface based registrations that take into account the sulcal and gyral anatomy also help mitigate these confounds, and may be the future direction of the field (Fischl, 2012; Van Essen, 2012).

On top of these issues, functional areal boundaries (see Parcellation Choice) of heteromodal association areas are variable with respect to sulcal and gyral anatomy (Fischl et al., 2008). New multimodal approaches that also anchor surface-based registrations to other structural or functional characteristics of an individual are likely to improve subject to subject registrations, and hence reduce intersubject variability, even more (e.g.(Glasser et al., 2016b; Robinson et al., 2014). Unfortunately, these advanced methods require more data acquisition time than is typically afforded for developmental studies. Nonetheless, a high quality anatomical image that is capable of being segmented properly into grey and white matter by itself is enough for the most basic suface-based registration tools.

Test-retest reliability

A final point regarding the use of rs-fcMRI for network mapping concerns its reliability as a measurement tool. Multiple studies have found evidence for limited test-retest reliability using a multitude of conventional rs-fcMRI network characteristics and analytic approaches (Braun et al., 2012; Wang et al., 2011; Wisner et al., 2013; Zuo et al., 2010; Zuo and Xing, 2014). Recent evidence suggests that these effects may largely result from insufficient data acquisition. For instance, FC networks in one highly sampled individual converged as a function of the timeseries length used to extract correlations and were extremely reproducible using approximately 1 hour of motion-free data (Laumann et al., 2015). However, this length of scan time for each individual is likely prohibitive in most developmental rs-fcMRI studies, underscoring the importance of internal study replication. In addition, more work could be done to develop analytic methods with both high sensitivity to individual variation and high test-retest reliability (see (Miranda-Dominguez et al., 2014a)).