The Seductive Allure of Brain Images
For the scientist (and for the general public), there is something especially captivating about colored images of the human brain, as if at last we could get to see directly “inside the black box.” But knowing that area X is activated (“lights up”) tells us nothing, unless it adds a new explanation or new prediction beyond the behavioral measures. If developmental neuroimaging is to have predictive and clinical power, we should be able to reverse engineer and go from brain activation to a better than chance guess as to which cognitive function is being used. As a tool, brain imaging is no better than a pencil or a fishing trip, unless it is hypothesis-driven. If the key goal of cognitive neuroscience were to find “simple and direct connections between brain and behavior” [Raizada et al., 2010], then phrenology could raise its ugly head again! But of course, as these authors go on to point out, the key goal never turns out to be simple and direct. Indeed, is it ever the case that invoking the maturational coming on-line of a specific brain region explains developmental change? In fact, most regions in the infant brain are partially active from very early on; what develops is best captured at the level of the complexities of changing intra- and inter-regional networks [Johnson,2001].
What approaches really bring new insights from brain imaging? Work in the clinical field certainly has. One example is the discovery that co-activation of frontal and limbic regions is predictive of patients' response to anti-depressant drugs, explaining why certain patients respond to drugs and others do not [Langenecker et al.,2007]. However, to my knowledge, there are no such studies in children of differential brain responses to drugs, such as the effects on the developing brain at different ages of the use of methylphenidate to treat ADHD. Such treatment/brain-response correlations can be extended to examining the usefulness of brain imaging at the cognitive level, for example, in understanding how/why intervention works or fails in different aged children [Posner and Rothbart,2005]. Neural data can also be particularly relevant when considering identical behavioral scores across typically and atypically developing populations [see also, Luna et al., this Special Issue, on the possibility that, even in the typically developing population, alternative circuitries may underlie equivalent behavior at different ages]. The demonstration that the brains of individuals with, say, Williams syndrome, reveal different neural processing in terms of localization and specialization of function from the brains of healthy controls when viewing faces [Karmiloff-Smith et al.,2004], despite equivalent behavioral scores on standardized face processing tasks, can yield new hypotheses about atypical brain development, leading to more informed intervention techniques as well as predictions about other aspects of the cognitive system that may have developed in a similarly deviant way [Karmiloff-Smith,2009]. Radically differing γ-band activity in two syndromes both purported to focus on featural detail [Grice et al.,2001] suggests that the cognitive notion of “featural” needs reconceptualizing. In general, cross-syndrome comorbidities at the behavioral level may be differentiated at the neural level, and if they are not, then we may have to rethink traditional diagnostic categories [Shaw et al., this Special Issue].
Interestingly, as the brain continues to mature in the healthy elderly—another developing group—language comprehension seems to be spared. It turns out that despite neuronal degeneration in the left cerebral areas active in younger controls during comprehension tasks, proficient comprehension in the ageing brain is sustained by compensatory bilateral frontotemporal and parietal activity, thereby offering a neural-level explanation for the behavioral level [Tyler et al.,2009]. Another example of the importance of examining the neural level comes from sensitivity to foreign language contrasts. It had long been thought that the capacity to discriminate non-native phonemes disappears towards the end of the first year of life. Indeed, although 6-month-old infants can distinguish non-native contrasts, this ability is absent behaviorally around 9–10 months when the infant comes to specialize in the phonemic repertoire of his/her native tongue. But neuroimaging of adults who show no behavioral signs of distinguishing the non-native contrasts reveals that their brains are in fact still registering the differences [Gaxiola-Rivera et al.,2000], highlighting the use of both the cognitive and neural levels of explanation and pointing to the potential for successful phonemic training even in the adult. And a developmental approach is critical to move the field away from simplistic notions like risk alleles, as Casey et al. (this Special Issue) make clear, because an allele that is a risk factor for one period of development or one type of environment may turn out to be a protective factor for others.
Methodological issues are rife in the area of neuroimaging. Although findings seem clear as brain images are flashed on a screen, the data reduction issues behind the scenes are still far from resolved. One concerns selection bias. The problem of automatically mapping an adult Region of Interest (ROI) directly to the child brain is discussed in the articles by Luna et al. and by Poldrack (this Special Issue). But just as serious is the possibility that the criteria used in the first place to define the adult ROI are the same as those being tested for in the data from that region, meaning that any correlations may be biased [Rajeev et al., 2010; see also discussion in Konrad and Eickhoff, this Special Issue]. In other words, is it the case that certain brain-behavior correlations actually only emerge when a set of specifically selected features is used? One recent solution to the ROI problem applied a single statistical test—linear regression—across the whole brain at once, without the need to carry out any pre-selection of voxels or features [Rajeev et al., 2010]. In other words, the study asked whether the degree to which there is separability of patterns on a whole-brain level between two conditions correlates with individual differences in a behavioral variable. Although ROI analyses may be useful if the targeted region is the right one, they do run the risk of selection bias. Rajeev and collaborators circumvented this risk by using a whole brain analysis and yet they successfully identified strong brain-behavior correlations for two distinct cognitive domains, despite having to deal with large numbers of uninformative voxels [Rajeev et al., 2010]. This method of analysis across the whole brain at once might be particularly useful when analyzing developmental data where brains change across time or for analyzing data from atypically developing brains where cerebral volumes/boundaries may differ substantially from the typical case.
Selection bias is not, of course, the only problem facing the field. Another revolves around the comparison of child and adult data if the two age groups are not performing at a similar level of accuracy and/or reaction time [see discussion in Church et al., this Special Issue]. A further problem lies in the different limitations of each of the imaging methods themselves. Non-invasive neuroimaging techniques suitable for children come in two forms: they either detect direct changes in the electrophysiological activity in the brain, such as electroencephalography (EEG) and event related potentials (ERP), or they detect the resultant hemodynamic changes in cerebral blood flow, such as functional magnetic resonance imaging (fMRI) and functional near infrared spectroscopy (fNIRS). Direct measures capture the millisecond temporal details of brain activity, whereas indirect methods are far more accurate in terms of spatial location. In adults, but not in children because the method is invasive, transcranial magnetic stimulation (TMS) can directly target a hypothesized brain region and then be used to support causal rather than merely correlational arguments [Cohen-Kadosh et al.,2007; Mevorach et al.,2010; see also, discussion in Galván, this Special Issue].
Curiously, when one speaks of “brain imaging,” it is often taken for granted that this only refers to MRI or fMRI and, with a few exceptions, magnetic resonance imaging is the focus of most of the articles in this Special Issue. Yet, fMRI is just one among several functional imaging methods available to developmental neuroscientists, and by no means the most child-friendly one. For example, much less space was given in the articles here to the huge body of developmental neuroimaging research using ERPs, and no mention was made at all of the very promising optical imaging method, fNIRS, whose design has been rapidly progressing for use with infants over the past decade. Moreover, only the Poldrack article (this Special Issue) highlighted the importance of computational modeling for understanding the neurobiology of developmental change, yet such models are critical for testing hypotheses and generating new predictions [Elman et al.,1996].
Let's briefly look at some of the advantages and limitations of each of the imaging methods in turn, without of course claiming to be exhaustive, since many of the articles in this Special Issue already focus in detail on the methodological problems facing developmental neuroimaging. Of course, of concern to all methods is the fact that rapid brain growth as well as changes in head shape/size during early ontogeny must be taken account of in any analyses that span developmental time.
One of the major advantages of fMRI, be it for developmental or adult neuroimaging, is its very fine spatial resolution. Another is that MRI—specifically diffusion MRI (DTI)—allows the scientist to measure developmental changes in gray and white matter, as discussed in several articles in this Special Issue [Paus; Konrad and Eickhoff; Blakemore et al.]. This is important because the two primary types of data available from DTI studies—water apparent diffusion coefficient and diffusion anisotropy measures—reveal significant developmental changes over time. Animal models demonstrate how DTI can discretely delineate the microstructure of white matter and gray matter in embryonic and early postnatal mouse brains [Zhang et al.,2006], highlighting the developmental potential of the approach.
The complex issues concerning the possible effects of developmental differences in vascular physiology and of interpreting negative BOLD signal change are thoroughly addressed by Church et al. (this Special Issue). One practical limitation of fMRI for use with infants and children is that participants need to remain very still during fMRI data acquisition, without which movement artefacts render the data unusable. For this reason, when researchers opt for fMRI as their method of choice, they have either focused on very young infants who cannot yet move or they sedate them to prevent them from moving, or the focus has been solely on older children. In sedated participants, other than bright flashes of light that tell us little about visual cognition, the tendency has been to test auditory discriminations which the brain processes during sleep. Furthermore, because of the slow data acquisition (several seconds, compared to milliseconds in ERP and fNIRS), the fMRI signal cannot distinguish anticipatory processing from stimulus-generated activity. This can be a problem when assessing, for instance, the extent to which there is a developmental shift from data-driven to top–down neural processing. Moreover, when comparing changing brain processes across age or when comparing atypical brains to typically developing ones, the BOLD signal cannot distinguish between moderate activation of a greater number of neurons versus increased activation within the same neurons. Such differences could be relevant to human developmental change, with animal models providing illumination on this question [Bandettini and Ungerleider,2001; Logothetis et al.,2001].
One of the major advantages of fNIRS systems over, say, fMRI, is that they are inexpensive and portable. They can also tolerate a degree of movement, which is critical when testing awake infants sitting upright on their parent's lap. Also, fNIRS is even more suitable for infants than for older children and adults because the optical geometry of the infant head renders biological tissue more transparent to light in the near infrared part of the spectrum (particularly in those with no hair!). Moreover, fNIRS can acquire data at a rapid temporal rate [Huppert et al.,2006,2008], overcoming some of the intrinsic limitations of fMRI mentioned above. fNIRS surpasses EEG in providing a better spatial resolution, thereby allowing more accurate localization of brain responses to specific cortical regions [for an excellent review of fNIRS over the past decade, see Lloyd-Fox et al.,2010]. fNIRS has been used to measure infant brain activity across a wide array of cognitive domains such as object processing [Wilcox et al., 2005], social communication [Grossmann et al.,2008], biological motion processing [Lloyd-Fox et al.,2009], action observation [Ahimada and Hiraki, 2006], and face processing [Blasi et al.,2007]. Another advantage is that fNIRS can distinguish between anticipatory and reactive neural processes.
Although fNIRS produces better spatial resolution than EEG and better temporal resolution than fMRI, its temporal resolution is lower than EEG and its spatial resolution not as good as fMRI. Currently, fNIRS sits between the advantages and limitations of the two methods, but that also constitutes its major advantage. Unlike MRI, NIRS does not lend itself to the measurement of brain structure [Minagawa-Kawai et al.,2007, 2009], so debates continue as to how to locate the origin of the hemodynamic response from the scalp measurements, whether there is in fact an infant-specific pattern of hemodynamic response, and how precisely to interpret that response [Aslin and Mehler,2005; Meek,2002]. Finally, the headgear is still very heavy compared to the ultra-light HD-ERP net, although work on improving the fNIRS methodology is actively underway [Everdell et al.,2005; Lloyd-Fox et al.,2010].
ERP provides an excellent recording of the temporal processes of neural activity. Initially, it was limited by the small number of sources on the original caps for children, but with the advent of the Geodesic net enabling high density ERPs (HD-ERP), the infant scalp is now fitted with 128 channels, thereby becoming one of the methods of choice for infant and child brain imaging. As with fNIRS, the method circumvents any fear of scanners in children. Indeed, the child can sit upright, alone, or on the parent's lap. Unlike fMRI, HD-ERP can be used to distinguish predictive responses from data-driven responses showing, for instance, how 9-month-old infants exhibit sub-threshold motor activity, i.e., sensorimotor alpha band suppression, while they observe the actions of others, that matches directly the neural signal occurring during their own actions [Southgate et al.,2009]. This points to the existence of a form of mirror neurons in the infant brain.
Although HD-ERP yields relatively good scalp maps, it obviously does not provide as good spatial location data as fMRI or fNIRS, so source reconstruction continues to pose a problem, particularly when the cognitive processes are complex and depend on large networks. Movement artefacts can also be more of an issue for ERP than fNIRS when acquiring child data.
The Need for a Multimethod Converging Approach
This very brief overview of the advantages and limitations of various direct and indirect methods of measuring brain activity makes it clear that no method alone will provide a full account of the developing brain. What is needed is simultaneous data acquisition using different methods which would lead to complementary, converging data about changes in the time course, the spatial location and connectivity of neural activity [Casey et al., this issue; see also Huppert et al.,2008; Steinbrink et al., 2006, for a multimodality approach with adults]. In infants, Grossmann and collaborators (2008) have used more than one method for the study of social perception of eye gaze, measuring fNIRS and EEG oscillatory signals in each of two different groups of infants. A small number of infants was tested with both systems, but this involved sequential rather than simultaneous multimodal acquisition of data, which must be a goal for future research.
Child Neuroimaging Versus Developmental Neuroimaging
I often start lectures with the provocative statement: Studying infants and children has little or nothing to do with development! This can hold for behavioral and genetic studies, and to a great extent also holds for brain imaging studies although, as mentioned, there are clear exceptions in the articles in this Special Issue [see also Casey and Durston,2006, for discussion]. Neurimaging of the developing brain involves a theoretical attitude of mind, a truly developmental framework, and a focus on progressive change [Karmiloff-Smith,1992, 1998]. Comparing the adult brain with that of a single, specific age group of children is not a developmental study; at best it constitutes a pilot study for a truly developmental approach to gradual change over time. Say, for a given cognitive function, the child brain activates region or network X, and the adult brain region or network Y. This indeed demonstrates that child and adult brains process stimuli differently, but it tells us nothing about the developing brain, i.e., the mechanisms underlying the developmental trajectory that brought the brain from X to Y. Although there are now a few studies which have compared children at more than one age to adults [Galván et al.,2006; Minagawa-Kawai et al., 2009; van Duijvenvoorde et al.,2008; Velanova et al.,2009, and articles in this Special Issue], the vast majority of albeit interesting work on the infant and child brain tend to focus on a single age group like 9-month-old or 5-year-old, or on the comparison of one age group of children with adults. Rare are those who ask truly developmental questions about how neural networks gradually change over time. Moreover, as Poldrack (this Special Issue) points out, there remains a substantial gulf between the wealth of neurobiological knowledge about early infant brain development, on the one hand, and the changes being examined in most neuroimaging studies in children over 5 years of age, on the other.
Many developmental questions remain to be fully addressed. Do the boundaries between regions change developmentally? If they do, then it is questionable to generalize adult Regions of Interest to the child brain. For a given cognitive function, are the same intra- and inter-regional networks active? This is unlikely, given the lengthy time required for white matter tract development [Poldrack, this Special Issue], together with significant differences in the ratio of gray to white matter in child and adult brains, as well as changes in plasticity over time [Huttenlocher,2002; see discussion by Galván, this Special Issue]. Is the only solution a (cost-intensive and labor-intensive) longitudinal study of the changing brain, as provided by the rich multisite, normative study of brain structure in children described in the article by Paus [this Special Issue; see also, Giedd et al.,1999]? Although longitudinal studies are not without problems [see discussion in Shaw et al., this Special Issue], this is indeed one data-rich source for studying the developing brain over time, but by no means the only one. First, detailed developmental trajectories have been built cross-sectionally at the cognitive level [Karmiloff-Smith et al.,2004; Thomas et al.,2009] which can then be verified on much smaller groups longitudinally, and there is no reason why the cross-sectional trajectory approach cannot be applied to developmental neuroimaging. Second, training studies over short periods can be very informative in adults and children [see discussion in Galván, this Special Issue]. A cleverly designed developmental study of cerebral change over a mere 2 days of training might tell us more about changes in adult and child brains across age than two single groups separated by many years of age but each measured only once. Adult number training studies in numerical problems and their neural consequences are already underway [Ansari,2008], and should be extended more often to development over time. Developmental approaches also lead to bidirectional questions: Instead of automatically concluding that, say, reduced parietal volume is the cause of numerical impairments, one can ask whether atypical processing over developmental time in neurons in parietal or different cortical areas caused a progressive reduction in parietal volume. Only a developmental approach can answer such questions.
Examples of Neural Changes Across Development
Typical fine tuning of brain functional organization is an activity-dependent process [Kandel et al.,2000]. According to Huttenlocher , plasticity changes over developmental time, with some mechanisms available throughout the lifetime (increase in synaptic strength, decrease in local inhibition, dendritric sprouting, formation of new synapses, formation of new neurons), whereas others only available to the early developing brain (use of unspecified labile synapses including silent synapses, competition for synaptic sites, persistence of normally transient connections, myelination). Changes in plasticity turn out to be region specific, suggesting [Thomas,2003] that there is no such general thing as “the brain's plasticity.” Many structures (e.g., dendrites, axons, and synapses) initially undergo exuberant growth followed by a period of pruning in which the processing of environmental input gradually sculpts the resulting brain structure.
Neural processing tends initially to be diffuse across several regions in both hemispheres, but with developmental time and the continuous processing of inputs, brain activity becomes increasingly restricted to more specific networks [Brown et al., 2008; Durston et al.,2006 and the discussion by Poldrack, this Special Issue, on the problem of the neurobiological plausibility of the gradual focalization hypothesis made on the basis of neuroimaging data]. If there is a gradual process of modularization over developmental time [Karmiloff-Smith,1992], as opposed to the notion of built-in modules, then this is likely to improve processing efficiency. A recent study by Minagawa-Kawai and colleagues  examined language-specific phonemic contrasts in infants from 3 months to 28 months and found differing age-specific onset of varying regions of the cortex. Another study suggests that comprehension of single words moves from bilateral processing between 13 and 17 months to left lateralized processing at 20 months [Mills et al.,1997]. Like vocabulary development, processing of human faces starts out with bilateral activity, with the brain displaying similar signatures for other stimuli like monkey faces [de Haan et al., 2002; Pascalis et al.,2001]. By the end of the first year, however, the brain becomes increasing fine-tuned for processing human faces, with other stimuli displaying different neural signatures, as well as increasing localization for human faces to specific networks in the right hemisphere [de Haan et al., 2002; Peelen et al.,2009]. We need to know more about how hemispheric differences influence developmental neural change. In adults, the RH seems to be involved in more parallel, coarse-grained, integrative processing, and the LH in more serial, fine-grained, predictive processing. How does this develop in children? Is information passage through corpus callosum always faster from RH to LH than from LH to RH, or does this alter over developmental time? Certainly, the thickness of the corpus callosum fibers changes developmentally over a lengthy period of time between infancy and adolescence [Keshavan et al., 2003]. Finally, short-range gray matter connectivity is greater in children, while white long-range connectivity develops considerably more slowly over time [Huttenlocher,2002]. All of these and other developmental changes must be taken into account when analyzing neuroimaging data over time. And, as shown in several articles in this Special Issue [Luna et al., Blakemore et al., Shaw et al.; see also, Crone et al.,2008], neural development doesn't end in childhood; puberty and adolescence witness many significant changes, which raises the important question of whether neural changes in learning and development are the same [Galván, this Special Issue]. Moreover, rather than focusing on chronological age in studies of adolescents, onset of puberty might be a more sensitive way of understanding interactions between hormonal and gray matter development, gender differences, as well as the general structural reorganization of the adolescent brain [Blakemore et al., this Special Issue].
Importance of Resting State Functional Connectivity
In adult neuroscience, there has been renewed interest in resting state fMRI (R-fMRI), focusing on a default circuit comprising a large network of brain regions, associated with task-irrelevant mental processes: precuneus/posterior cingulate cortex, medial prefrontal cortex and medial, lateral, and inferior parietal cortex. Studies point to a high degree of functional connectivity during rest, i.e., inter-regional temporal synchrony [Raichle et al., 2001; see also discussions in Luna et al., Konrad and Eickhoff, this Special Issue]. As self-organizing processes are a necessary part of the explanation of how the brain changes over developmental time, it is critical to understand whether and how the default mode network changes developmentally [Konrad and Eickhoff, this Special Issue].
Spontaneous brain activity during sleep, for instance, plays a critical role in the consolidation of memory across the lifespan, involving redistribution of memory representations from temporary hippocampal storage to neocortical long-term storage sites. The dialog between neocortex and hippocampus generates sharp-wave ripples and is orchestrated by the <1 Hz EEG slow oscillation during slow wave sleep. Unlike adults whose sleep patterns involve cycles of slow wave sleep to REM sleep, young infants fall directly into REM sleep, with the proportion of slow wave sleep increasing only slowly over developmental time [Hill et al.,2007]. How this affects the neural processes involved in sleep-related consolidation of learning in children remains to be fully elucidated.
The importance of developmental changes in resting state brain activity will, I believe, become very prominent in the next few years. Many years ago [Karmiloff-Smith,1992] I put forward a cognitive-level developmental hypothesis—the Representational Redescription Hypothesis—postulating that what is specifically human to human intelligence is a process by which task-specific representations stored in the brain become, via an internally generated process of representational redescription, domain-general knowledge to the brain. This internal self-organizing process, I argued, is generated by behavioral mastery, not by negative feedback, and allows knowledge relevant to one domain to become transportable to other domains without the need to process new external input. With the current advances in brain imaging, it should be possible to assess the hypothesis by detecting specific networks in cerebral resting state underlying representational redescription.
Futuristic? An Idealized World for Developmental Brain Imaging
Let us end on some speculations as to where the field might go. Clearly multiple converging methods on the same individuals will be paramount in providing increasingly detailed accounts of spatio-temporal neural network changes over developmental time. Subtle individual differences in brain development will, in my view, take centre stage, instead of being considered noise to be controlled for in large group studies [see discussion in Paus, this Special Issue, on the emergence of population neuroscience as a field at the intersection of cognitive neuroscience, genetics, and epidemiology]. A combination of in-depth longitudinal case studies with cross-sectional brain trajectory approaches is critical, combined with a deeper understanding of acceleration and deceleration of developmental trajectories of brain development [Shaw et al., this Special Issue]. More focus is required on differences in regional connectivity rather than identifying specific regions, and this may be particularly important when examining atypical development [Konrad and Eickhoff, this Special Issue]. And, as neuroscience and genetics become increasingly sophisticated, we will need to examine how polymorphisms and their downstream effects at the genetic level [including the surprising amount of allele-specific methylation, Schalkwyk et al.,2010] might relate to subtle differences in brain structure or function, as well as elucidating more in-depth phenotypics and refining our approach to environmental influences [Paus, this Special Issue]. Gene expression will need to be viewed as dynamic, rather than static, in different environmental and developmental contexts [Casey et al. and Paus, this Special Issue]. And animal models must compare similar task demands at the cognitive level when examining the phenotypic effects of identical cross-species genes [Karmiloff-Smith,2009].
The use of simple tasks, with children lying flat in a scanner, may no longer suffice. Studies will need to get out of the scanner and into the real world, using neuroimaging methods that allow the child to move naturally around the environment, thereby capturing every change in attention [see discussion in Paus, this Special Issue]. After all, each time the child brain takes a step forward, the environment dynamically changes too. Perhaps in the not too distant future, we will witness the invention of a pushchair with two built-in imaging devices (one superior spatially, the other temporally), together with a portable eye tracker, to measure subtle changes in both brain activity and attention. And later, as the child becomes mobile, the neuroimaging devices and eye tracker could be inserted into a T-shirt or cap. This is not so far-fetched when one recalls just how recently neuroimaging technologies have been introduced into the lab. If we are really to understand progressive changes in the human brain, we have to take a multilevel dynamic approach to epigenetics, brain structure/function, network connectivity, cognition, behavior, and the environment, in which the notion “developing” really counts.