The future of functionally-related structural change assessment

Abstract

The brain is continually changing its function and structure in response to changing environmental demands. Magnetic resonance imaging (MRI) methods can be used to repeatedly scan the same individuals over time and in this way have provided powerful tools for assessing such brain change. Functional MRI has provided important insights into changes that occur with learning or recovery but this review will focus on the complementary information that can be provided by structural MRI methods. Structural methods have been powerful in indicating when and where changes occur in both gray and white matter with learning and recovery. However, the measures that we derive from structural MRI are typically ambiguous in biological terms. An important future challenge is to develop methods that will allow us to determine precisely what has changed.

The brain is not a static structure but instead changes its function and structure over time in response to environmental or intrinsic factors. Such changes can be studied in exquisite detail in animal models, but the advent of neuroimaging brought about possibilities for studying these processes in humans. The non-invasive nature of blood oxygen level dependent functional MRI (BOLD fMRI) was particularly exciting in this regard as it opened the door for taking repeated functional measurements in the same individuals over time. This is in contrast to positron emission tomography (PET), for example, in which potential risks of repeated exposure to radiation limited possibilities for longitudinal studies. Not only functional, but also structural changes occur with learning, development and recovery from brain injury. Multiple non-functional MRI modalities offer opportunities to study these phenomena and it could be argued that there are certain advantages to testing for structural, rather than functional, measures of brain changes. In this review, I will discuss ways in which functional, structural and diffusion MRI have been used to study functionally-relevant brain changes.

Use of functional MRI to study experience-dependent brain plasticity

As someone learns a new task, FMRI can identify short-term changes in the brain regions involved in carrying out that task (Doyon et al., 2002; Floyer-Lea and Matthews, 2004) and computational modeling of BOLD signal changes can determine more precisely the computations performed by different elements in these circuits (Brovelli et al., 2008; Grafton et al., 2008). The ability to scan the same individual repeatedly over days to weeks has also allowed for longer-term changes to be characterized. Such longer-term learning and consolidation is largely associated with decreasing activity, though specific regions (such as parts of cerebellum, putamen and motor cortex) can show increases at later stages of learning in relation to acquisition of specific aspects of the new motor skill (Floyer-Lea and Matthews, 2005; Puttemans et al., 2005; Steele and Penhune, 2010).

Improvement in performance with repeated practice is also a key concept in recovery of function after brain damage such as stroke (Krakauer, 2006). Rehabilitation can be thought of as relearning of movements that have been impaired as a result of stroke. Interestingly, the effects of practice on brain activity measured using FMRI seem to differ markedly between healthy people and patients after stroke. While practice in healthy people is predominantly associated with decreasing activity as discussed above, rehabilitation in chronic stroke patients can be characterized by increasing recruitment (Enzinger et al., 2009; Johansen-Berg et al., 2002; Whitall et al., 2011). This apparent difference holds up when effects of motor practice are directly contrasted between groups: two weeks practice of a motor task was associated with decreasing activity in healthy people and increasing activity in patients (Bosnell et al., 2011). Many of the areas showing increasing activity over time in patients were also found to have impaired structural connectivity at baseline, suggesting that structurally compromised regions can, to some extent, be functionally ‘rescued’ through practice (Bosnell et al., 2011). The effects of practice on activity in the chronic phase after stroke are quite distinct from changes in activity over time with recovery in the first few months post-stroke, when reductions of initial overactivations are typically found (Buma et al., 2010; Feydy et al., 2002; Ward et al., 2003).

Limitations of FMRI in studying brain change

Although FMRI has shed important light on experience-dependent changes in the brain, there are limitations to the use of FMRI for longitudinal studies of learning or recovery. Task-based BOLD FMRI signals from a given individual scanned repeatedly over time are highly variable (McGonigle et al., 2000; Zandbelt et al., 2008). Quantitative estimates of reproducibility differ a great deal, dependent on methodology, subject group, and measures used, but coefficients of variation greater than 50% are not uncommon (Kimberley et al., 2008; Marshall et al., 2004). This means that possibilities for detecting subtle genuine effects on top of this background of between-session, within-subject noise are limited. Alternative functional measures, such as perfusion measures derived from arterial spin labeling, may have greater stability over time and so may be better suited to studies aiming to detect low frequency changes in functional responses (Wang et al., 2003). In contrast to some FMRI measures, structural MRI measures can have reasonable test-retest reliability, with coefficients of variation <7.5% often reported (Han et al., 2006; Heiervang et al., 2006; Wang et al., 2008), and so might be more sensitive to experience-dependent longitudinal change.

For studies of recovery, the suitability of FMRI depends on the patient population under investigation. BOLD FMRI is potentially problematic in the study of patients with vascular disease as typical assumptions about neurovascular coupling may not hold (Krainik et al., 2005) and so the BOLD response may not accurately reflect underlying neuronal activity. For example, in some stroke patients clear neuronal responses can be detected using non-vascular methods such as magnetoencephalography despite the absence of a detectable BOLD response (Rossini et al., 2004).

Structural brain change: another route to detecting functionally relevant brain plasticity?

Animal studies show that learning is not only characterised by changes in functional responses but also by physical restructuring in brain tissue. For example, growth of new neurons occurs in the adult hippocampus in response to learning or to exercise (Gould et al., 1999) (although the existence of cortical adult neurogenesis is controversial (Rakic, 2002)). Other types of both neuronal and non-neuronal experience-dependent structural plasticity are consistently found. For example, activity or learning is associated with synaptogenesis (Kleim et al., 1996), angiogenesis (Adkins et al., 2006), increase in glial cell size and number (Kleim et al., 2007; Li et al., 2005), and cortico-cortical rewiring (Hihara et al., 2006). Structural change can happen rapidly; optical imaging reveals increased spine formation within an hour of rats learning a new reaching task (Xu et al., 2009).

So animal studies convincingly show that brain structure is continually shifting in response to learning and environmental change. But the structures that are changing are tiny compared to the scale of our imaging voxels. To what extent are experience-dependent changes in brain structure detectable by MRI?

Structural MRI can detect functionally-relevant brain change

One of the first suggestions that grey matter measures derived from T1-weighted structural MRI are sensitive to experience came from Maguire and colleague’s seminal studies in London taxi drivers. They showed that taxi drivers, with their encyclopedic knowledge of London’s notoriously complex street plan, had enlarged posterior hippocampi (Maguire et al., 2000). The possibility that this structural difference was a result of experience (rather than a pre-existing brain characteristic that compelled people to become taxi drivers) was supported by a further finding that the size of this brain structure correlated with the number of years spent driving taxis. Even stronger evidence that experience shapes adult human brain structure has come from longitudinal studies, in which the same individuals are scanned serially, before and after some targeted training. So, for example, learning to juggle is associated with grey matter increases in the visual motion areas (Draganski et al., 2004) that can be detected after as little as a week of training (Driemeyer et al., 2008). Training regimes do not have to be physical – cramming for exams (Draganski et al., 2006), learning Morse code (Schmidt-Wilcke et al., 2010), mirror reading (Ilg et al., 2008) and practicing mindfulness (Holzel et al., 2011) have all been shown to increase grey matter in specific regions.

Another route to detecting functionally-relevant brain change is using MR spectroscopy (MRS), a method for non-invasive quantification of neurochemicals from within a volume of interest. One measurable metabolite is N-acetylaspartate (NAA), a chemical that is found within neurons and whose concentration is generally assumed to reflect neuronal structural and functional integrity. In one study in which people were trained on a spatial navigation task over 4 months, increases in NAA were detected in the hippocampus, potentially reflecting increases in neuronal or glial size or number, or altered metabolism (Lovden et al., 2011).

Experience-dependent structural changes are not limited to grey matter; white matter also appears to be susceptible to such effects. Diffusion tensor imaging (DTI) provides measures that reflect white matter microstructure. For example, fractional anisotropy (FA), quantifies the directional dependence of water diffusion in tissue and, in white matter, it is modulated by tissue characteristics such as membrane integrity, myelin, axon diameter, packing density, and geometry (Beaulieu, 2009). Bengtsson and colleagues first showed that pianists had higher FA than non-pianists in structures including the corticospinal tract (Bengtsson et al., 2005), but, more interestingly from the perspective of experience-dependent brain change, they also showed that FA in this pathway correlated with the number of hours spent practicing during early childhood. The amount of time spent practicing during adolescence and adulthood correlated with FA in other pathways including the corpus callosum and arcuate fasciculus. These results suggest that white matter is susceptible to experience-dependent structural modification and that different pathways have different sensitive periods during development. However, it could be argued that pre-existing factors could have influenced both brain structure and people’s propensity to practice and so, again, it is useful to look to longitudinal studies for more definitive evidence that experience shapes white matter structure.

Longitudinal studies of WM structure show that learning to juggle results not only in GM increases in occipito-parietal cortical areas involved in reaching and grasping, but also in FA increase in underlying white matter pathways (Scholz et al., 2009). Other interventions that have been shown to evoke changes in WM structure in adult brain include balance training (Taubert et al., 2010) and memory training (Engvig et al., 2011; Takeuchi et al., 2010).

How plausible are the reported changes?

Reports of changes in gross brain structure after a behavioural intervention of a few weeks have sometimes been met with skepticism. How believable are these observations? Have they been replicated? Are the size and time course of effects plausible?

The general observation that measures of brain microstructure show experience-dependent change has now been made across multiple laboratories and in many different training contexts (Draganski et al., 2004; Draganski et al., 2006; Driemeyer et al., 2008; Engvig et al., 2011; Holzel et al., 2011; Ilg et al., 2008; Keller and Just, 2009; Schmidt-Wilcke et al., 2010; Scholz et al., 2009; Takeuchi et al., 2010; Taubert et al., 2010). I am aware of only a single study reporting a negative result (Thomas et al., 2009): Thomas and colleagues found that 2 weeks training on a motor adaptation task did not produce any changes in grey matter measures. The fact that only one negative study exists could be taken as good evidence for the plausibility of the phenomenon, or could simply reflect publication bias; perhaps far more negative results have been observed but not reported.

Another important observation of the Thomas study was that if the scans were co-registered in a biased (but commonly conducted) way, by using a single timepoint (e.g,. the baseline scan) as a target to which other timepoints are aligned, then spurious structural changes were detected. When scans were registered appropriately (by defining an unbiased target, positioned at a mid-point between timepoints), no effects were found. It is important to clarify whether effects previously reported using ‘biased’ registration would still stand if appropriate registration was employed, and also that future longitudinal studies are sure to use the non-biased approach.

There have been replications of specific patterns of structural change within a given behavioural training context. However, perhaps unsurprisingly, particular labs have tended to focus on a specific training regime and so it is often the case that a given regime has only been studied in a small number of labs. So, for example, the taxi driver model has been studied by Maguire and colleagues in London whereas the majority of juggling studies have been conducted by May and colleagues in Regensburg. Good replication of effects has been achieved within these research groups, though details can vary. For example, for grey matter change with juggling training, the Regensburg group consistently report changes in an area they describe as the human homologue of V5 (Draganski et al., 2004; Driemeyer et al., 2008), though it is not always found bilaterally (Boyke et al., 2008) and is accompanied by changes in other areas that vary from study to study. Replication across groups has been less frequently attempted. A juggling study from my own group (Scholz et al., 2009) found grey matter changes in parietal cortex, 2-3cm medial to parietal changes reported in some previous jugging studies (Draganski et al., 2004; Draganski et al., 2006) but was unable to replicate the previously reported V5 effect, even when using the same methodology as the prior studies. There are always differences in the precise timing and delivery of training, and also in the approaches to image analysis, that could explain these failures to fully replicate effects across studies and it will be important for future studies to rigorously test the possibilities. It is also worth pointing out that failure to replicate is an all too common phenomenon (Lehrer, 2010).

Another consideration relevant for assessing the plausibility of observed changes is the magnitude of the effects. When effect sizes have been reported for longitudinal studies of change in structural or diffusion measures in healthy adults they are typically in the order of 1-5% for grey matter (Draganski et al., 2004; Draganski et al., 2006; Driemeyer et al., 2008; Holzel et al., 2011; Ilg et al., 2008; Taubert et al., 2010) and a similar range (0.5%-5%) for white matter (Scholz et al., 2009; Takeuchi et al., 2010; Taubert et al., 2010). These effects appear broadly plausible when compared to what is observed in clinical populations or model organisms, where much larger biological changes would be expected. So, for example, in patients with multiple sclerosis, FA is reduced by 10-20% compared to controls in so-called ‘normal appearing’ white matter, where subtle neurodegeneration is thought to occur (Roosendaal et al., 2009); in post-mortem brain samples, FA within lesions, where degeneration, demyelination and inflammation may all occur, is around 50% lower than in normal appearing white matter (Schmierer et al., 2008). In model systems, it is possible to selectively determine the effects of a given biological feature on FA. So, for example, anisotropy is reduced by around 20% in myelin deficient rat spinal cord (Gulani et al., 2001).

Although the reported effect sizes are in a biologically plausible range, some questions could be raised over the direction of change, particular for measures such as fractional anisotropy, where the biological interpretation is complex. For example, while most training studies reports increase in FA with experience, one study found that balance training led to increases in GM but to decreases in FA over time (Taubert et al., 2010). This highlights the challenge of interpreting biologically ambiguous imaging measures (Zatorre et al., 2011). While FA increases could be speculated to reflect activity-dependent increase in myelination or fiber density, FA decreases could be observed if factors such as increases in axon diameter, or maturation of a secondary fiber population, dominate.

Finally, the time scale of changes can be a source of skepticism. Is it plausible that the brain changes after just a few weeks, or even days (Driemeyer et al., 2008), of training? Animal studies suggest that structural change, at least at a cellular level, can happen very quickly. The number of spines in rat motor cortex, for example, increases within one hour of practicing a novel motor skill (Xu et al., 2009).

Methodological considerations

Studies of experience-dependent structural change can make use of a wide variety of imaging measures, analysed with a range of different approaches. So, for example, changes in grey matter can be assessed using voxel-based, deformation-based or tensor-based morphometry (Good et al., 2001), or using cortical thickness measures (Fischl and Dale, 2000). Changes in white matter can be assessed using diffusion tensor measures assessed voxel-wise, using a tract skeleton (Smith et al., 2006), or using regions of interest, which can be defined on the diffusion maps or by using tractography. A full discussion of the available options is beyond the scope of the current review but it is worth underlining that differences in the choice of technique, and also in critical analysis steps such as image registration (Thomas et al., 2009) or spatial smoothing (Jones et al., 2005), can significantly impact on results.

Studies also differ in the number of timepoints considered. In addition to scans acquired before and immediately after training, some have conducted follow up scans to assess the persistence of structural change (Draganski et al., 2004; Draganski et al., 2006; Scholz et al., 2009), while others have added timepoints during training, to assess how rapidly effects can be detected (Driemeyer et al., 2008). If the aim is simply to assess with maximal accuracy the magnitude of training related effects (rather than their timecourse), then it can be informative to look to the clinical trials literature, where the question of optimal timing of measurement points has been carefully considered (Winkens et al., 2005). Such considerations might, for example, suggest that additional measurements are best added at the end of an intervention, rather than spaced throughout it.

The choice of appropriate controls is an important consideration in designing plasticity studies. Longitudinal FMRI studies of learning effects will typically include a control task condition and test for a time by condition interaction to demonstrate that training-related changes are specific to the trained task. This is a particularly important consideration in longitudinal FMRI studies as systematic effects of time are often found, at least if two time points are considered; less activation is commonly detected the second time a task is performed, presumably due to habitation-like effects (Goodyear and Douglas, 2009). The situation is not quite the same with longitudinal studies of brain structure where, at least over time scales of a few weeks, systematic time-dependent effects (e.g., atrophy/development) would not be expected in a healthy cohort. It could, therefore, be argued that a control group is unnecessary in such studies. However, including a control group that does not receive training provides a control for at least theoretically possible effects such as changes in scanner performance over time.

What are we measuring?

MRI studies have been a useful complement to animal studies in furthering our understanding of how experience shapes brain structure. Particular advantages of imaging in this regard include the ability to study humans, opening possibilities for studying functions such as language, and for investigating patient populations. MRI also typically provides us with measures from across the whole brain, rather than needing to focus on specific pre-defined areas, and is very feasible to carry out at multiple time points in the same individual. Imaging has therefore been highly valuable in showing when and where brain structure changes in response to experience.

However, there are also clear limitations to imaging measures when compared to what could be assessed in animal studies. The spatial resolution of imaging measurements is far cruder than what can be achieved for most animal studies; structural changes would need to be of a sufficient magnitude and spatial extent to result in a bulk effect within an imaging voxel. When an effect is large enough to be detected using imaging, although that provides a powerful demonstration that something is changing in a particular location in the brain, it is not able to tell us precisely what is changing. When an increase in gray matter is found, this could reflect change in neuronal size or number or increased dendritic arborisation or synaptogenesis. It could also reflect non-neuronal factors such as glial change or angiogenesis. Similarly, when a white matter change is found, such as an increase in FA, this could reflect increasing axon number, increased myelination, or even decreased axon diameter, or decreasing strength of a crossing fiber population. We have an increasingly good understanding of how changes in any given tissue property might influence our imaging measures, but making the reverse inference is extremely difficult – faced with a change in an imaging measure it is typically not possible to infer which tissue property has changed as there is not a one-to-one relationship between most imaging measures and underlying biological structures (Zatorre et al., 2011).

How should we interpret structural brain change with MRI?

There are a number of approaches that are being taken that will help us to interpret structural change more accurately in future. First, an empirical approach is to test for relationships between change in MRI measures taken from animals and change in microstructural properties measured using conventional techniques such as histology. For example, Lerch and colleagues trained mice on the Morris water maze, a well-characterised spatial navigation task, known to depend on the hippocampus (Lerch et al., 2011). Structural MRI revealed an increase in hippocampal volume in trained animals, whose brains were then processed for histology. Stains for angiogenesis and glial cells were included but the only stain that correlated with the MRI volume measures was a stain for growth-associated protein-43, a marker for neuronal process remodeling. Blumenfeld-Katzir and colleagues trained rats on the same maze task and then acquired diffusion MRI scans (Blumenfeld-Katzir et al., 2011). They found increased FA and decreased mean diffusivity in both grey and white matter regions in the trained animals compared to control animals. Histological measures were used to determine underlying mechanisms. The dentate gyrus, which showed reduced diffusivity with imaging, was found to have increased staining for markers of synapses and astrocytes. This contrasts somewhat to the Lerch study in which glial markers did not correlate with grey matter change (Lerch et al., 2011). However the different imaging measures considered by the two studies could conceivably have different biological correlates. The studies also differed in the time elapsed between training and scanning and therefore may have been sensitive to different biological processes. In addition to assessing gray matter change, the study by Blumenfeld-Katzir and colleagues also considered changes in white matter and found that the corpus callosum, which showed increased FA with imaging, was found to have increased staining for myelin (Blumenfeld-Katzir et al., 2011).

Another approach to gaining greater biological insight into how experience shapes brain structure is to use imaging measures with greater specificity for the underlying biology. For example, measures taken from magnetization transfer imaging, or from relaxometry, are argued to reflect myelin content specifically (Laule et al., 2007), and susceptibility-weighted imaging has particular sensitivity to iron content (Haacke et al., 2004). Such measures could complement measures derived other modalities with differential sensitivity to underlying change. Analysis approaches which explicitly use information from multiple modalities to mutually inform estimation of parameters of interest could be used to maximize the benefits of multi-modal acquisition (Woolrich et al., 2009).

Finally, recent efforts at micro-structural modeling aim to create informed models of tissue properties and acquire data optimized to estimate these properties. For example, a model that assesses diffusion of water within restricted tissue compartments can be applied to diffusion imaging data to provide estimates of axon diameter distributions (Assaf et al., 2008) and variants on that model can be used to estimate axon diameter and density in human brain even in voxels containing multiple fibre populations (Alexander et al., 2010; Zhang et al., 2011).

Conclusions

The advent of functional MRI opened up new possibilities for studying the same individual over time in order to gain more sensitive insights into learning, development and recovery. Structural MRI modalities can complement this functional information by providing measures sensitive to physical remodeling in tissue. With all these techniques one limitation is uncertainty over precisely what is being measured. While structural measures are powerful for telling us where and when something has changed, they cannot currently provide definitive answers as to what precisely has changed. Future efforts to integrate information across modalities, and to develop more sophisticated microstructural models, should help to maximize the utility of MRI for assessing functionally-relevant brain change.