The notion that functional connectivity at rest can be closely related to grey matter structure has recently emerged¹. Importantly, a few studies investigating patterns in healthy brain structure have provided mechanistic insights into the spread or selective targeting of disease processes^1-3. To establish whether grey matter structural covariance patterns underlie the full repertoire of functional networks used by the brain at ‘‘rest” would provide fundamental insights into the nature of such functional networks. Another key question is whether these grey matter volume networks might be associated with variation in cortical folding, or cortical thickness⁴. To uncover the richness of structural patterns in the human brain, we have used here a purely data-driven approach to co-model three complementary types of grey matter information – grey matter volume, thickness and area – on a large, healthy population covering most of the lifespan (n=484, aged 8-85y, 220 males, right-handed).

* All 484 healthy volunteers (Oslo) underwent the same imaging protocol on a 1.5T Siemens Avanto scanner, with no hardware upgrades and only minor software upgrades (12-channel head coil, MPRAGE T1w, 1.25×1.25×1.2 mm³, 2 repeats).

* We obtained voxel-wise GM volume by processing T1w images using FSL-VBM⁵, as well as vertex-wise cortical thickness and surface area measures using FreeSurfer⁶. Using FLICA⁷, a linked ICA approach implemented in FSL⁸, we obtained an automatic decomposition of the images into 70 independent components (IC) to make it comparable with Smith et al.⁹. Each of these spatial components represents a mode of variation of brain structure across all 484 healthy participants.

The majority of the structural components (~40 of the 70 ICs) recapitulated functionally-meaningful brain networks observed using fMRI at rest and using tasks⁹ (Fig. 1). None of these components describing regional modes of variation were significantly associated with age, gender or ICV. The richness of the functional networks described by these structural ICs extended beyond the canonical resting-state networks. We found a meaningful split between the two DMNs (Fig. 1), representing the contribution of the limbic and associative parts of the precuneus¹⁰, and identified the visual functional network of the precuneus¹⁰ (Fig. 3). In addition, we singled out in the structural ICs the same 8 visual components presented in Smith et al.⁹ (Fig. 2). We also found a notable separation in the cerebellum between the mainly motor lobules (I-VI), the lobules VII and VIIIa – with a fine segmentation of the Crus I and Crus II –, and lobules VIIIb and IX including the corresponding vermis (Fig. 2). Interestingly, each of these assemblies of cerebellar lobules shares common functional connectivity with the cerebral networks¹¹.

Next, using hierarchical clustering with the correlation coefficient of subjects weights as similarity measure, we investigated whether these fine-grained structural networks would further cluster together to form functionally relevant higher-level networks. The ICs strongly clustering together were (all in Fig. 3): i) an auditory- with a language-related IC; ii) “Visual 1” with “Visual 3”, further clustering with the pair of “Visual 2” and “Visual 7”; iii) notably, “DMN 2” with prefrontal regions and IPS, forming a structural substrate for most of the DMN and its negatively correlated functional network¹² (Fig. 4); iv) “DMN 1” with “Executive” to create a cluster comparable with the limbic precuneus functional network¹⁰ (Fig. 4); v) the left and right frontal and parietal ICs to explicitly make the lateralised fronto-parietal networks; vi) “Cerebellar 1” with “Cerebellar 2”, and “Cerebellar 3” with anterior temporal pole areas and “Cerebellar 4”; vii) a cluster comprising of the structural equivalent of the temporal FFA (“Visual 6”), posterior STS and PPA (“Visual 5”), and one cluster made of the EBA (“Visual 4”) and the occipital FFA.

One further benefit of using such a linked-ICA approach was its multi-modal aspect which revealed that the modes of variation in GM volume across the 484 healthy participants spatially co-varied with cortical area, i.e. differences in folding (e.g., Fig. 5). GM volume networks involving primary sensory areas such as V1, M1, S1 or A1 also partially co-varied with cortical thickness (e.g., Fig. 5).

We found that modes of variation of GM volume across a large number of healthy subjects reflect the entire repertoire of functionally-meaningful architecture that can be observed within-subject using fMRI. These regional structural networks are not associated with age and are mostly related to variation in cortical area. These results demonstrate the anatomical nature of functional networks at a population level and suggest that “neurotrophic” events must occur during development (and evolution?) to impose a common folding pattern of distant brain regions across subjects.