Cortical thickness gradients in structural hierarchies

1Although elegant work has established an indisputable general relationship between cortical morphology, cytoarchitecture and function, the precise nature of this relationship is unclear. Moreover, there is a considerable degree of inter-individual variability in the large-scale structural features of the brain, which adds further ambiguity to regional analysis of both structure and function. The goal of the current study was to determine whether one widely used morphological measure—cortical thickness—varies across the cortex in a manner predicted by other cytoarchitectural and functional characteristics. Specifically, we used existing knowledge about the cellular, laminar and hodological patterns within the cortex to determine whether cortical thickness measures followed a corresponding pattern. Such an observation would lend weight to the proposition that macroscopic structural measures obtained using MRI, act as a useful marker of underlying neurophysiology. If true, then individual differences in cortical thickness measures would offer a possible interpretational framework for the neural bases of psychiatric disorders.

2Central to our investigation is the principle that the brain processes information through pathways known as functional hierarchies. Computational hierarchical models describe a system of functionally specialised regions with feedforward and feedback connections. Feedforward connections communicate incoming sensory information, while feedback connections relay experience-derived predictions, which alter the response to the incoming signal. Each cortical region processes particular features of the incoming sensory stimulus. These features tend to become increasingly specific and complex on ascending the hierarchy. For example, the primary visual cortex, V1, is responsive to basic image features present in most stimuli, whereas area MT is preferentially responsive to certain types of motion and the fusiform face area (FFA) is selective to faces, suggesting MT and FFA are higher than V1 in the visual processing hierarchy. Similar patterns of increasing stimulus selectivity and complexity are found in auditory and somatosensory hierarchies.

3The cortex also demonstrates hierarchical organisation of regions based on structural characteristics, such as cortico-cortical connectivity. These data are most commonly derived through histological tracer studies in which dyes are injected to follow efferent axons to their target areas. The relative positions of two cortical areas within in a hierarchy are determined by the ratio of connection types, which can be feedforward (generally from lower to higher regions) or feedback (higher to lower). Feedforward axons tend to project from more superficial cortical layers while feedback axons project from deeper layers. Tracer studies have revealed connection-based hierarchies in visual, somatosensory, motor, auditory and prefrontal regions.

4At a microscopic scale, the cytoarchitecture of cortical areas mirrors their structural hierarchical organisation. The degree of laminar differentiation is, for example, strongly related to a region's connectivity and position within a sensory structural hierarchy. Primary sensory regions have six clearly defined cytoarchitectural layers whereas, for higher-level areas, the laminar differentiation is less distinct. More generally, it has been observed that primary sensory regions have increased neuronal density compared to other cortical areas.

5While laminar structure and density might reflect the relative position of a region in a structural hierarchy, these patterns cannot yet be quantified using MRI. This limits the degree to which human structural and functional processing can be correlated in vivo. However, it may be that other more accessible parameters such as cortical thickness are related to cortical cytoarchitecture and can be used as an effective surrogate marker of laminar differentiation and, by extension, structural hierarchy. For example, neuronal density and cortical thickness are generally inversely correlated, and both exhibit a similar rostral-caudal gradient across the cortex. These results tentatively support the hypothesis that a gradient of cortical thickness, from thinner primary sensory areas to thicker higher-order areas, may be a useful biomarker of changes in cortical cytoarchitecture, structure and potentially functions of the hierarchy.

6While there is some debate over the precise ordering of individual regions, structural hierarchies do show consistent similarities to functional hierarchies. For example, in the visual cortex, area MT is consistently higher than areas V1 and V2; in the auditory cortex, there is a matching progression from the auditory core, to the belt regions, and then to the superior temporal gyrus and superior temporal sulcus; and in the somatosensory cortex, the hierarchy progresses caudally across the post-central gyrus, through BA (Brodmann area) 3a, 3b, 1, 2 and 5. Structural and functional hierarchies may therefore describe the same principle of cortical organisation. However, several electrophysiological studies have demonstrated higher-order characteristics in lower order regions, suggesting that sensory processing is not so easily simplified. Nevertheless, if structural hierarchies index functional hierarchies, then inter-individual variability in cortical cytoarchitecture and thickness may reflect differences in hierarchical function.

7However, empirical evidence that cortical thickness is a marker of structural hierarchy is lacking. A relationship between the thickness of a cortical area and its relative hierarchical position may be obscured by the mechanics of cortical gyrification, which cause gyri to be on average 20% thicker than sulci. However, these effects, although relatively under-explored, can easily be identified and accounted for using MR-based surface reconstruction approaches (Fig. 3).

8In summary, this study began by evaluating whether cortical thickness is a surrogate marker of laminar differentiation. Subsequently, we investigated whether connectivity and functional hierarchies are organised in accordance with structural gradients. We used structural MR images from macaque and human subjects. For each subject, thickness gradients were quantified for three separate hierarchies in each species, namely, visual, somatosensory and auditory. Hierarchical level was identified in the macaque using standardised parcellation maps and tracer-derived hierarchies. With respect to the human data set, given a lack of precise knowledge about the layout of structural hierarchies, we used geodesic distance from the primary sensory region as a proxy measure of hierarchical position, having validated this approach against macaque hierarchies (Fig. 1). Geodesic distance also offers a means of analysis which obviates differences between the various hierarchical models that exist. Finally, we compared thickness against a functionally derived visual hierarchy. Importantly, all measures of cortical thickness were adjusted to account for the local effects of folding.

9Structural MRI data from one macaque (Macacca mulatta) and 83 human subjects were analysed. The macaque data were acquired using a 3 T Acheiva Phillips Scanner at Boston University School of Medicine. A T1-weighted 3DMPRAGE image (TR = 7.09, TE = 6, Flip = 8) was obtained through the brain with 0.6 x 0.6 x 0.6 mm voxel size. The human subjects were scanned according to the Human Connectome Project protocol, using 3 T Siemens Skyra Connectome Scanner at Washington University, St. Louis. A T1-weighted 3DMPRAGE image (TR = 2400, TE = 2.14, TI = 1000, Flip = 8, FOV = 224 × 224) was obtained through the brain with 0.7 x 0.7 x 0.7 mm voxel size.

10Cortical reconstructions were created using FreeSurfer software. Cortical thickness was measured as the shortest distance between each vertex on the white matter surface and the pial surface.

11The macaque surface reconstruction was registered to the F99 atlas using FreeSurfer & Caret software. The F99 atlas contains several cortical parcellation schemes including FVE91, used to outline somatosensory and visual cortical areas, PHT00 used for auditory areas and M132 atlas used for comparison with laminar differentiation data.

12Macaque visual and somatosensory hierarchies were taken from Felleman and Van Essen; regions were delineated in the FVE91 parcellation scheme. Successive levels in the hierarchy were given numerical values, such that V1 has a hierarchical level of 1, V2 is level 2, etc. Several areas in the original visual hierarchy were omitted from this analysis: 7b, ER, 36, HC, FEF, 46, TF and TH. The first 4 of these are non-visual regions, TF and TH part of the parahippocampal cortex, which is structurally different to normal neocortex. Areas FEF and 46 are frontal regions and not part of the hypothesised continuous structural gradient. The somatosensory hierarchy was simplified from the somatomotor hierarchy (for discussion of motor hierarchy, see Limitations section).

13The macaque auditory hierarchy is described by several studies, which outline hierarchies with consistent characteristics, including anterolateral progression. For this study, we took the hierarchy from, a cytoarchitecture and tracer-defined hierarchy with 6 levels. Architectonic regions in the hierarchy are well described and each level is an aggregate of several areas in the PHT00 parcellation scheme. In this study, the auditory hierarchy was manually outlined into these 6 cortical regions.

14Human reconstructions were registered to the PALS-B12 atlas in Caret and into its component gyri and sulci in FreeSurfer.

15The human visual cortex was parcellated with a matching functional hierarchy, adapted from using regions found in the PALS-B12 atlas and V1 as identified by FreeSurfer. Primary somatosensory cortex (BA 3b) (parcellated in FreeSurfer; ), post-central gyrus and sulcus, and superior parietal gyrus were included as somatosensory regions. Primary auditory cortex (Heschel's gyrus/transverse temporal gyrus), superior temporal gyrus and transverse temporal sulcus were included as auditory regions.

16Due to an absence of matching functional hierarchy and parcellation schemes, geodesic distance from the primary sensory regions was used as a surrogate marker for hierarchical level for somatosensory and auditory hierarchies.

17There is some debate over the precise details of cortical hierarchies, but there is evidence that, within the visual hierarchy, hierarchical level increases with distance from the primary sensory region. For this reason, we used geodesic distance from the primary sensory region to each randomly parcellated target region as a surrogate measure for hierarchical level. Geodesic distance was measured as the shortest path across the white matter surface between the vertices closest to the centre of each region (Fig. 3). Distance was first validated against hierarchical level and cortical thickness in the all three macaque sensory hierarchies and human functional visual hierarchy, before being implemented as the sole measure of hierarchical level in human auditory and somatosensory hierarchies. We acknowledge that the use of the geodesic distance as a marker for hierarchical level was necessarily speculative in human somatosensory and auditory cortex but, in the absence of an empirically validated means of identifying levels, we adopted these measures as a reasonable surrogate.

18Automated parcellation schemes are to some extent limited by individual variability. For example, there is a two-fold intersubject variability in the surface area of V1, and many borders are not consistently associated with large-scale morphological landmarks. To address the uncertainty over precise border locations, the cortex was randomly parcellated into 100 regions of approximately equal surface area; any region containing a border between different hierarchical levels was assigned the average value of the levels (Fig. 2a). By repeating random parcellation 10 times and averaging the thicknesses (Fig. 2b), we minimise the effect of folding (Fig. 3) and the bias of each individual random parcellation scheme.

19Because high cortical myelin content, as found in the primary sensory regions, can lead to underestimations in some MR measurements of cortical thickness, we validated FreeSurfer cortical thickness in the human auditory cortex by comparing it to previously published histological data. Average FreeSurfer values for human auditory cortex were 2.84 mm and 2.88 mm for the left and right hemispheres (SD = 0.31 mm), while histological measurement gives an estimate of 2.9 mm. The close match with histological data is in keeping previous studies, which have validated FreeSurfer thickness measurements with histological comparison.

20Measurements of laminar differentiation in the macaque cortex were derived from an aggregate of multiple histological studies. Brain areas in the M132 atlas were ranked according to their cytoarchitectonic differentiation, taking into account features such as neuronal density and granularisation of layer 4. For example, the primary visual cortex with strong laminar differentiation has laminar differentiation type 8, while areas with less clearly defined laminar structure have progressively lower rankings.

21Statistical analysis was carried out using Matlab & R. All relationships were tested using Spearman's partial rank correlation controlling for the effect of hemisphere. We also tested macaque data using a linear model, with individual hemispheres as covariates and human data were tested with a linear mixed-effects model, with individual hemispheres as covariates and individual subjects included as a random effect.

22Regional laminar differentiation data were compared with folding-corrected, MRI-derived cortical thickness values using Spearman's partial rank correlation controlling for the effect of hemisphere. In agreement with our hypothesis, cortical thickness correlated negatively with laminar differentiation type (rs = − 0.39, p < 0.001)(Fig. 4).

23Macaque data were analysed using Spearman's partial rank correlation and a linear model, controlling for the effect of hemisphere (see Table 1). In agreement with our hypothesis, hierarchical level was significantly correlated with and predicted by cortical thickness in visual, somatosensory and auditory cortices, where hierarchical level increased with thickness (Figs. 5, 8 & 9). In the somatosensory cortex, the correlation was not as strong but was still significant.

24Hierarchical level was significantly correlated with and predicted by geodesic distance in all three hierarchies. This result is important for human somatosensory and auditory hierarchies, where there was no available parcellation for the literature-based hierarchies.

25Finally, geodesic distance from the primary sensory region was also predicted by cortical thickness in the visual, somatosensory and auditory cortices. Again the relationship in the somatosensory cortex was not as strong, but the correlation was still significant.

26Human data were analysed using Spearman's partial rank correlation and a linear mixed-effects model, controlling for the effect of hemisphere and the random effects of individuals (see Table 1). In agreement with the macaque data and our original hypothesis geodesic distance was significantly predicted by cortical thickness in the visual, somatosensory and auditory cortices (Figs. 6, 8 and 9). For somatosensory and auditory cortices the correlations were not as strong but were still significant (Fig. 9).

27For the visual cortex, functional hierarchical level was significantly predicted by cortical thickness and geodesic distance from the primary visual cortex (Figs. 7 & 8).

28In this investigation, we tested the hypothesis that cortical thickness may be a useful marker of cytoarchitecture—one that can be used to identify gradients in sensory structural hierarchies. Our results demonstrate a strong relationship between cortical thickness, laminar differentiation and estimated hierarchical position for visual, auditory and somatosensory hierarchies. Furthermore, our data are compatible with a close coordination of cortical structure and function. Critically, the findings overall suggest that patterns of cortical thickness, which are readily measurable in vivo with structural neuroimaging, can be interpreted meaningfully in terms of underlying patterns that relate directly to cortical function. They offer a framework for analysing and interpreting cortical thickness measures in health and disease.

29In the macaque, cortical thickness—corrected for effect of folding—was negatively correlated with laminar differentiation. While previous studies have shown a relationship between thickness and neuronal density, an inverse correlation with laminar differentiation has not previously been demonstrated. Moreover, throughout the cortex, there are a number of inter-dependent and predictable relationships between various cytoarchitectural properties. For example, neuronal density can be reliably related to intracortical connectivity, such that lower neuronal density is associated with increased dendritic arborisation and density of synapses. The inter-relationship between neuronal density, laminar differentiation and cortical thickness may be useful in considering thickness changes in studies of structural MRI.

30Cortical hierarchies have previously been characterised by laminar differentiation and by layer-specific cortico-cortical connections. Here we have shown that cortical thickness may be a further indicator of hierarchical level. Each of these measures brings its own advantages and functional implications.

31First, layer-specific interregional connections are considered functionally as indexing feedforward (signals going from lower to higher regions) and feedback connections (higher to lower signalling). Hierarchical connectivity patterns such as these have previously been characterised using axonal tracer studies indicating that (i) feedforward connections predominantly originate in upper cortical layers and terminate in the lower layers of a target region and (ii) feedback connections project from lower and terminating in upper layers.

32Interregional feedforward and feedback connectivity has been related to integration of sensory information in the context of increasingly influential predictive coding models. Within such models, the balance of feedforward/feedback connections determines functional organisation. Thus, the patterns of laminar-specific connectivity have important implications for how structure may govern and shape function.

33A second measure of structural hierarchy is cytoarchitectural. The pattern of feedforward/feedback innervation is closely linked to laminar differentiation, neuronal density and even dendritic tree size and spine density. Furthermore, some of the changes in neuronal density are layer specific, where caudal cortical regions have increased density in supragranular layers, which give rise to feedforward axons. Rather than describing an identical, repeated neuronal microcircuit, these measures emphasise systematic variation in intracortical connectivity throughout the hierarchies.

34Our present study shows that cortical thickness follows a similar gradient to hierarchies of connectivity and cytoarchitecture and might therefore offer a third related marker of structural hierarchy. Critically, unlike laminar connectivity and cytoarchitectural patters, thickness gradients are accessible to standard neuroimaging tools. Thus, the identification of this non-invasive, in vivo marker of cytoarchitectural trends offers a framework for interpreting cortical thickness patterns which are frequently reported, although poorly understood, in health and disease. Moreover, our observation of a gradient of cortical thickness within the fMRI-based human visual hierarchy suggests a hitherto unestablished mapping between structural and functional hierarchies. Although such a relationship remains speculative, our results support the idea that structural regularities have a direct functional correlate.

35Laminar connectivity, cytoarchitecture and thickness change over the course of development. There is little change in the number of feedforward connections postnatally, while both feedback connections, synaptic density and cortical thickness increase postnatally up to a peak and then decrease. Based on this rationale, we hypothesise developmental changes in thickness gradients mark changes cortical cytoarchitecture and connectivity that support higher-level sensory processing.

36During childhood, thickness increases at a differential rate across the cortex, with various regions achieving peak thickness at different time points between the ages of 7 and 11. Of note, hierarchical gradients develop in a gradual and non-uniform way, with primary sensory areas reaching peak thickness before secondary and higher association areas. It has been observed that the timing of thickness peaks is correlated with functional development. For example, in the visual system, the occipital cortex peaks at roughly the same age that children develop letter acuity and global motion detection. These observations suggest that by quantifying the thickness gradient in a sensory structural hierarchy, it may be possible to more accurately interpret the stages of development.

37More generally, a gradient-based approach to structural analysis obviates another important limitation of such analysis, namely, the high inter-individual variation in the pattern of morphology and cytoarchitecture across the cortex. By assessing gradient differences across structural hierarchies as opposed to individual regions, the ambiguity of regional identity is somewhat mitigated. Moreover, larger-scale regional analyses might be informative in conditions of atypical neurodevelopment such as autism and schizophrenia, which are associated with complex distributed changes across the cortex and are not reducible to a single morphological or functional abnormality.

38It is important to note that observations of the relationships between morphology, cytoarchitecture and functional organisation were made having taken steps to account for the effect of folding, which causes gyri to be significantly thicker than sulci (Fig. 3b). This folding effect may obscure the large-scale thickness gradient of interest, and we therefore minimised its effect through the random parcellation-based smoothing (Figs. 2, 3c and d). In fact, this approach is almost equivalent to surface-based Gaussian kernel smoothing.

39It is noteworthy that, in general, individual differences in cortical folding patterns could obscure or introduce group cortical thickness differences. Smoothing at a study-appropriate scale may help to mitigate these folding effects. Indeed this might underpin increases in sensitivity to group thickness differences observed when applying surface-based Gaussian smoothing kernels at 30 mm FWHM.

40A number of limitations attend this study. First, there is an ongoing debate over the precise organisation of structural hierarchies and it can be difficult to assert that any given sub-region belongs to a single specific level within a hierarchy. Moreover, it is simplistic to suppose that structural hierarchies comprehensively and exhaustively describe functional organisation and information flow, and there are several possible measures of hierarchical function. We aimed to mitigate these concerns by demonstrating only the general trend of cortical thickness in structural hierarchies. Having done this in those hierarchies about which we were most confident, a complementary analysis using geodesic distance as a surrogate marker of hierarchical levels was used to further validate the approach. Although this approach overlooks some of the subtleties of structural hierarchies, we nevertheless believe our results persuasively demonstrate a variability in cortical thickness that closely correlates with previously established markers of structural hierarchy and certain measures of functional hierarchy.

41In this study, we also confirmed the validity of our measure of cortical thickness, given the confounding effect of cortical myelin content (see Methods). High cortical myelin content, as found in the primary sensory regions, can lead to underestimations in some MR measurements of cortical thickness. However, our analysis of the auditory cortical thickness indicated a close match between FreeSurfer cortical thickness measurement and histological data, which is in keeping previous validation studies.

42Some of our results in the somatosensory and auditory cortices, while significant, are not particularly strong (rs < 0.35). This may in part be due to the hierarchies being relatively small, both in terms of the number of hierarchical levels and the area over which they extend. Nevertheless, the results remain significant suggesting that these structural principles do still apply in these sensory areas.

43We note that this analysis was confined to the cortical thickness gradients in sensory processing hierarchies, such that these results may not be generalisable to motor, frontal and prefrontal cortices. While there is evidence for the existence of motor and frontal hierarchies, both structural and functional, several considerations prevent the simple extension of the hypothesis to include these pathways for this experiment. First, the structure of frontal and prefrontal cortices is far more variable. Unlike sensory cortices, primary motor and premotor areas have an agranular cytoarchitecture and are particularly thick, despite having clear laminar differentiation. Similarly, the frontal cortex does not follow the same inverse relationship between neuronal density and cortical thickness seen in all other cortical regions. Moreover, the direction of information flow is reversed, with signals predominantly travelling caudally from frontal ‘abstract’ areas towards the primary motor cortex. Therefore, while cortical thickness gradients may be of interest in frontal-motor hierarchies, they are unlikely to follow the pattern seen in sensory cortices.

44Finally, it should be noted that this experiment used data from a single macaque as a proof-of-concept analysis, capitalising on the high degree of specificity with which the functional organisation is known in this species. Although this necessarily limits the generalisability of this analysis, the closely matched findings in the 83 human subjects support the hypothesis that gradients of cortical thickness co-vary with hierarchical level and suggest that this may be a general principle of brain organisation and structure.

45Our findings offer a new framework for the interpreting cortical thickness measured using structural MRI. Cortical thickness gradients are significantly correlated with structural hierarchical level of sensory processing hierarchies in macaque and human subjects. Multiple lines of evidence suggest that this relationship is further characterised by predictable changes in cortical cytoarchitecture and connectivity. The results also suggest a close coupling between cortical structure and functional demand, such that higher-order visual areas tend to be thicker. We propose that analysis of these gradients may advance our understanding of the structure–function relationship in normal and atypical neurodevelopment.