Introduction
For many brain functions, including the processing of sensory information or the encoding of bodily movements, the two hemispheres are an almost perfect mirror image of each other, with the left perceiving and controlling the right and vice versa. However, there is ample evidence for several cognitive modalities to mainly operate in one of the two hemispheres. Although left dominant speech production and hand preference are arguably the most important lateralized feature of the brain (Broca
1861; Knecht
2000), there are others including face perception (Yovel et al.
2008; Wilkinson et al.
2009) and spatial attention allocation (Ciçek et al.
2009) that are thought to have a bias for one of the hemispheres. Detailed knowledge of how the neuronal architecture gives rise to asymmetries in distributions of functions can help our general understanding of how functions emanate from brain mechanisms. In addition, it may eventually shed light on more fundamental questions addressing hemispheric specialization regarding human conscious experience (Gazzaniga
2000).
Furthermore, investigating intersubject variation in dissimilarity of the hemispheric architectures may be important for revealing neuronal underpinnings of behavioral differences across subjects. While the theory of left and right brain personality should be regarded with considerable skepticism (Nielsen et al.
2013), this does not refute the existence and relevance of other variations in patterns of brain asymmetry across individuals, that relate to specific behavioral, physiological, or personality features. For instance, depression has been linked to an imbalance in activity between the hemispheres (Henriques and Davidson
1991; Flor-Henry et al.
2004; Nielsen et al.
2013), while incomplete lateralization is thought to be one of the neuronal abnormalities underlying schizophrenia (Stephane et al.
2001; Frith
2005).
Advanced neuroimaging techniques are now available that allow for detailed descriptions of the brain’s connectivity in terms of asymmetries and hemispheric specialization (Hervé et al.
2013). In this study, we investigate the neuronal underpinnings of lateralization of brain functions and intersubject variation therein by mapping hemispheric asymmetries of the connectivity structure. For investigating these asymmetries, we chose fMRI Resting-State (RS) activation, the spatiotemporally linked spontaneous fluctuations in the hemodynamic response in the absence of sensory input or motor activity (Biswal et al.
1997; Raichle and Mintun
2006; Friston
2011). While we acknowledge uncertainty in what kind of neuronal processes these fluctuations correspond to, and that correlations in RS fluctuations do not necessarily correspond to anatomical connectivity (Sporns
2011), they do contain important signatures of the underlying connectivity (Skudlarski et al.
2008; Raemaekers et al.
2014). Furthermore, fMRI RS is relatively easy to acquire and share, and it addresses the whole brain instead of tapping into specific functions or areas.
So far, fMRI RS studies have suggested a highly symmetric connectivity, with networks as detected with, for instance, independent component analysis being identically distributed across the two hemispheres (Damoiseaux et al.
2006; Smith et al.
2009). Although two frontoparietal networks have been consistently detected that each reside in one hemisphere only, they seem to be each other’s mirror image. While they exhibit different features with respect to their activity, the connectivity strengths within their respective hemispheres may thus be indistinguishable. Any asymmetries are, therefore, prone to be relatively minute. A direct quantification of the extent of the hemispheric symmetry is, however, missing. The first objective of this study is to measure the level of asymmetric connectivity by estimating how well the brain’s connectivity structure can be predicted by the connectivity structure of the brain mirrored over the longitudinal fissure.
Recent investigations have demonstrated the existence of hemispheric asymmetries in resting-state connectivity, and that the intersubject variation therein is related to language lateralization and handedness (Wang et al.
2014; Joliot et al.
2016). While asymmetries most likely involve the frontoparietal, attention, and default mode networks, no study so far has precisely mapped the exact underlying connectivity asymmetries in full. Studies so far have either focused on asymmetries of connections within the left hemisphere or the right hemisphere (Tzourio-Mazoyer et al.
2015), address specific regions of interest (Fox et al.
2006; Xiao et al.
2016; Hasler et al.
2017), or have obtained condensed asymmetry metrics by quantifying the mean amount of asymmetric connections per voxel (Joliot et al.
2016). While these approaches certainly have their merits, they do not provide a full and simultaneously assessment of any connectivity asymmetries between and within hemispheres. This endeavor is necessary to obtain a comprehensive overview on which connections might be underlying lateralized brain functions. In addition to the full assessment of asymmetries in individual connections, we apply a data driven approach to establish if there are other sources besides language lateralization that are driving individual differences in the observed asymmetries. As we anticipate small effect sizes regarding asymmetries, data of prolonged RS measurements are required in large groups of healthy volunteers, which are made available by the Human Connectome Project (HCP) (Van Essen et al.
2012b). The need for prolonged measurements also means that at this stage, we are disregarding potential dynamics in asymmetry. As metric we use Asymmetric Functional Connectivity (AFC), which we define as the extent to which the RS correlation matrices differ from those of the same brains mirrored along the longitudinal fissure. This represents a whole-brain survey of asymmetries in inter and intrahemispheric connectivity. As this metric is to a large extent explorative, we are at the current stage not employing graph-theoretical metrics. We classified patterns of asymmetry which were subsequently linked to fMRI language lateralization and a collection of behavioral features.
Discussion
According to our metric, the mean connectivity during RS was more than 95% symmetric. We found several coherent and consistent asymmetries nonetheless. The group-mean AFC consisted foremost of (1) higher correlations between language areas in the left hemisphere than between their right hemisphere homologues, and (2) higher correlations between the default mode network in the left hemisphere and language homologue areas in the right hemisphere, than between language areas in the left hemisphere and the default mode network in the right hemisphere. The extent to which individual subjects exhibited this pattern correlated with LTL and handedness. Further exploration in intersubject variation in AFC revealed several additional asymmetries, one involving entire hemispheres, and another involving correlations with limbic areas.
The anatomical distribution of resting-state networks already suggested a high level of symmetry, and this study concretely estimates this level at 95%. This implies that effect sizes of asymmetries are modest at best, meaning that prolonged RS measurements are necessary to obtain stable estimates. Alternatively, the high level of symmetry suggests that using the mirror connections in stroke or tumor research can, in many cases, be a valid strategy.
The mean AFC indicates stronger connectivity amongst language areas as compared to amongst their contralateral homologues, which agrees with language function as one of the key lateralized features of the brain. This finding also concurs with the presence of left lateralized hubs in left hemisphere language areas during RS as was detected using a graph-theoretical approach (Nielsen et al.
2013). In addition to the asymmetries in connections within hemispheres, this study also addressed asymmetry of between-hemisphere connections. Surprisingly, these included the strongest effects which involved the interhemispheric correlations between language areas and the default mode network, with stronger correlations between right language homologue areas and left default mode network, than between their contralateral homologues. While one might have predicted language areas to interact with the default mode network, it is rather unexpected that such interaction is lateralized. This observation is, however, in line with previously observed grey matter as well as RS asymmetries in the default mode network (Saenger et al.
2012).
Note that the asymmetry in the between-hemisphere correlations for language areas and the default mode network are by no means certain to originate from asymmetries in direct white matter tracts, considering the relatively sparse heterotopic as opposed to homotopic connectivity through the Corpus Callosum (Jarbo et al.
2012). However, we currently have no conclusive alternative model explaining how this asymmetry may have arisen. While theoretically it might be that this pattern is somehow the result of intrahemispheric inhibitory white matter connections (Singh and Fawcett
2008), our findings did not show pronounced asymmetries in intrahemispheric correlations between language areas and the default mode network. This would imply that such intrahemispheric inhibitory connections would have little direct influence on BOLD connectivity, which seems unlikely. The possible role of white matter asymmetries in our findings needs further exploration, perhaps using diffusion tensor imaging.
The extent to which individual subjects exhibited the pattern of mean AFC was predictive of language lateralization, as was the language component score. Several previous studies have demonstrated a relationship between aspects of resting-state activity on one hand, and language lateralization or handedness on the other (Wang et al.
2014; Tzourio-Mazoyer et al.
2015; Joliot et al.
2016). Correlates with language or handedness include the strength of homotopic (between homologue areas) connectivity (Tzourio-Mazoyer et al.
2015), the ratio of the strength of ipsilateral and contralateral connections of voxels (Wang et al.
2014), and the patterns of asymmetries of within-hemisphere connections (Joliot et al.
2016). The latter measure comes closest to the metric employed here, but did not include asymmetries of between-hemisphere connections. The current study elaborates on these previous findings by linking language and handedness to more specific asymmetric connections, including the aforementioned interhemispheric connections between language areas, and interhemispheric connections between language areas and the default mode network.
The strength of the correlation between AFC-score and LTL-score was only moderate (
r = 0.455), which might be caused by combined imperfect reliability of both measures. Whereas AFC-scores have good (but not perfect) reliability, the relatively short length of the language processing task (± 7.5 min) could have made the reliability of individual estimates of LTL suboptimal. A stronger relationship might have been observed if a longer language task were used. In addition, the original intent for the HCP language task was mapping semantic language processing, as opposed to more general language function as is common for the purpose of determining language lateralization. Performance of mathematical operations was used as reference condition, which may, by itself, be left lateralized (Burbaud et al.
1995; Krueger et al.
2008). The effect sizes of the LTL-scores may thus have been suboptimal, which might have attenuated the strength of the relationship between AFC-scores and LTL-scores. Nevertheless, the imperfect reliability of AFC-score even with the current state-of-the-art RS data prevents its use for determining language lateralization in individual subjects such as for presurgical mapping. The asymmetry does, however, provide a potential straightforward research tool for investigating hemispheric dominance in prelingual, as well as non-human subjects.
Handedness was the only metric of all anatomical, behavioral, and individual difference variables that had an at least moderate correlation with AFC (
r = 0.322). Although it is generally assumed that language lateralization was predictive of hand preference (Knecht
2000), recent findings have questioned this relationship when disregarding subjects who are strongly right language lateralized (Mazoyer et al.
2014). Although our study was not specifically aimed at investigating handedness and included far more right than left-handers, we observed the relationship with handedness in spite of the absence of strongly negative AFC-scores or LTL-scores in our sample (Fig.
6), which would represent an equivalent of strongly right language lateralized subjects. In theory, one might explain this discrepancy of results by arguing that handedness as a sensorimotor phenomenon, instead of language lateralization, determines the mean AFC matrix, and thus AFC-scores. We believe this to be highly unlikely, however, considering that all correlations between LTL-score, handedness, and AFC-score were significantly positive, and that asymmetries were not overrepresented in connections with motor areas as would be predicted based on the fact that hand preference can be decoded from RS correlations between the different motor areas (Pool et al.
2015). We have as yet no explanation for the apparent contradiction in results.
The relationship between AFC-score (the extent to which the group mean is represented in individual subjects) and task language lateralization shows that language lateralization to a large extent determines the group mean, most likely because the side of the language dominant hemisphere was highly unevenly distributed within the experimental population (78% were right handed). However, there are more sources that drive AFC in individual subjects. Using PCA, we found that intersubject variation in AFC was composed of several coexisting patterns, confirming the previous findings that asymmetry in RS is not a single source phenomenon (Liu et al.
2009). The moderate reproducibility of some of the principal component scores suggests that they most likely represent individual variation in the tendency to express particular asymmetric brain states or events (Petridou et al.
2013) instead of being hard wired asymmetries.
The sources explaining the most variance in the current sample were the hemisphere component and limbic component. The hemisphere component reflected relatively enhanced connectivity within either one of the two hemispheres, in combination with the contralateral Cerebellum (Fig.
7a). This component seems to match to the recently observed differences in the amplitudes of the global/mean signal variations between the hemispheres (McAvoy et al.
2015). Between-hemisphere differences in the amplitudes of global signal variations may arise when BOLD responses in one of the two hemispheres are more synchronized than in the other, which is exactly the type of asymmetry that this component is representing. The limbic component consisted of an asymmetry in connectivity between limbic and cortical areas, both between and within hemispheres. None of these components showed a significant relationship with any of the behavioral, anatomical, or individual difference measures that were acquired by the HCP. However, future studies may attempt to establish links with more elaborate behavioral measures that are assumed to have stronger links with hemispheric asymmetries. Such links may include personality measures such as ability for divergent critical thinking (Moore et al.
2009; Santarnecchi et al.
2015), but also pathological conditions such as depression and schizophrenia (Henriques and Davidson
1991; Stephane et al.
2001; Flor-Henry et al.
2004; Frith
2005; Moore et al.
2009; Nielsen et al.
2013; Santarnecchi et al.
2015).
The method for determining the ROIs plays a key role within the approach used here. We choose a Freesurfer automatic parcellation scheme (Fischl
2004; Desikan et al.
2006), which uses geometric information derived from the individual cortical model in addition to neuroanatomical convention. First, the amount of asymmetry is inversely proportional to how well the parcellation algorithm performs at defining two anatomically homologue areas, as AFC will increase when homologue areas are ill defined. Although errors in ROI definitions must have occurred as no perfect methods for areal segmentation exist, such errors cannot account for the current observations. While substantial random errors in segmentation would have prevented us from observing any effects, systematic errors can be refuted by a simple thought experiment. If two homologue areas are defined incorrectly, or more incorrectly than others, this would affect all connections involving the homologue ROI pair, not only specific ones. Anatomical asymmetry would be reflected by crossing vertical and horizontal lines of large effects (both positive and negative) in the asymmetry matrices. None of our findings meet this criterion. Second, the choice for these ROIs roughly determines the minimum spatial scale at which the asymmetries can be detected, which more or less matches the spatial scales of the RS networks as they are found by independent component analysis (Damoiseaux et al.
2006). While more detailed asymmetries could in theory be investigated using smaller ROIs, this would increase the amount of error in the ROI definitions, and reduce the statistical power of the study.
A major potential shortcoming of this study is confounding of AFC by large-scale susceptibility effects. It has been shown that such artifacts can lead to incorrect conclusions regarding lateralization of brain functions when the phase-encoding direction is either LR or RL (Mathiak et al.
2012). These effects were counteracted in the HCP pipeline by corrections for spatial distortions (Glasser et al.
2013), and by combining data sets with LR and RL phase-encoding directions. By testing asymmetry in the effect for LR vs. LR phase-encoding directions, we found that these countermeasures were almost fully successful, except for some connections involving areas close to the nasal cavities. Perhaps, around this area, even small systematic asymmetries in anatomical structure may lead to detectable asymmetries in large-scale susceptibility effects. However, none of the detected patterns of asymmetry primarily involved connectivity with these areas.