Abstract
Microstructural mechanisms underlying apparent cortical thinning during childhood development are unknown. Using functional, quantitative, and diffusion magnetic resonance imaging in children and adults, we tested if tissue growth (lower T1 relaxation time and mean diffusivity (MD)) or pruning (higher T1 and MD) underlies cortical thinning in ventral temporal cortex (VTC). After age 5, T1 and MD decreased in mid and deep cortex of functionally-defined regions in lateral VTC, and in their adjacent white matter. T1 and MD decreases were (i) consistent with tissue growth related to myelin proliferation, which we verified with adult postmortem histology and (ii) correlated with apparent cortical thinning. Thus, contrary to prevailing theories, cortical tissue does not thin during childhood, it becomes more myelinated, shifting the gray-white matter boundary deeper into cortex. As tissue growth is prominent in regions with protracted functional development, our data suggest an intriguing hypothesis that functional development and myelination are interlinked.
Acknowledgements
This research was funded by NIH grants 1RO1EY 02231801A1 and 1R01EY02391501A1 to KGS, 5T32EY020485 to VSN, and NRSA F31EY027201 to JG. We thank Prof. Dr. Thomas Arendt from the Paul Flechsig Institute of Brain Research, University of Leipzig, for providing post mortem brain tissue samples. The project was supported by the European Research Council under the European Union’s Seventh Framework Programme (FP7/2007-2013) / ERC grant agreement n° 616905, and the BMBF (01EW1711A & B) in the framework of ERA-NET NEURON.
As the brain develops from infancy to adulthood, magnetic resonance imaging (MRI) shows that cortex thins1-6. Early sensory regions thin before higher-level frontal and temporal regions1,3. However, the mechanisms underlying cortical thinning during development are not well understood.
Three developmental theories have been proposed to explain cortical thinning across development: pruning, myelination, and cortical morphology. Pruning, evaluated using cross-sectional histological studies on post mortem brains, suggests that the removal of inefficient synapses, dendrites, and neurons leads to cortical tissue loss7-9. This subsequently produces thinner cortex in adulthood. Pruning is thought to improve neural processing by optimizing brain circuits for particular operations. Myelination suggests that the myelin sheath wrapping axons grows during development2, 10-13, thus increasing the efficiency of saltatory conduction, and leading to faster and more reliable information transmission. As myelin increases raise the intensity of voxels in T1-weighted anatomical images, voxels close to the gray-white matter boundary will appear brighter in adults than children, thus shifting the gray-white boundary deeper into cortex in adults14. Cortical morphology (cortical folding and surface area) suggests that during development sulcification and surface area increase due to mechanical forces15,16, thereby producing thinner cortex in adulthood17. These mechanisms are not mutually exclusive as a combination of pruning, increased myelination, and morphological alterations may result in thinner cortex in adulthood.
Advances in quantitative MRI (qMRI)18-20 and diffusion MRI (dMRI)21 provide independent and complementary non-invasive, measurements of gray and white matter tissue properties, allowing us to disambiguate developmental hypotheses in the living human brain. Quantitative MRI enables the measurement and comparison across individuals of the amount of non-water tissue within a voxel (macromolecular tissue volume, MTV) and the relaxation time (T1), which depend on tissue composition (e.g., tissue containing myelin, reduces T1 more than tissue without it). Additionally, mean diffusivity (MD), obtained from dMRI, depends on the size, density, and structure of the space within tissue through which water diffuses and provides additional insight into microstructural changes during development20,22.
How can these MRI measurements differentiate the three developmental hypotheses? Although we cannot measure microstructure directly using in vivo MRI, we can distinguish these theories because their effects on microstructure within a voxel differs. Pruning results in developmental reductions in synaptic spines7, dendrites, and neurons. Although, the magnitude of the effect of pruning on T1 or MD remains unknown, pruning results in less cortical tissue for protons to exchange energy with and less hindrance on diffusion of water molecules, predicting higher T1 and MD in adult’s cortex than children’s. In contrast, myelination predicts developmental changes to both white and gray matter. In the white matter, increased myelination predicts lower T118,23 and reduced MD20,22. Myelination is most pronounced in deep cortical layers closer to gray-white boundary, which contains more afferent and efferent projections, than layers close to pial surface24, 25. Thus, in gray matter, developmental myelination predicts lower T126,27 and lower MD in adult’s than children’s cortex, especially in deeper layers. Finally, developmental changes in cortical morphology predict no development of either T1 or MD of gray or white matter. Instead it predicts morphological changes in the local cortical curvature and surface area.
We applied these novel methods to address developmental theories of cortical thinning. We obtained multiple independent measures of fMRI, qMRI, and dMRI in 27 children (ages 5-9, N=12, 9 female; ages 10-12, N=15, 5 female) and 30 adults (ages 22-28, 11 female). We used fMRI to define functional regions of interest (fROIs) selective to faces, characters, and places within ventral temporal cortex (VTC) of each participant using a functional localizer experiment (Fig. S1a). We focused on these VTC regions for two reasons. First, accumulating evidence shows differential functional28-31 and anatomical development26 across VTC, whereby face- and character-selective regions show prolonged development compared to place-selective regions. Thus, examination of the development of cortical thickness (CT) within VTC allows testing if thinning is guided by uniform or heterogeneous mechanisms across a cortical expanse. Second, functional regions within VTC can be identified in each participant’s brain without spatial smoothing, and have a reliable arrangement with respect to the cortical folding32, 33. Thus, studying VTC offers enhanced precision for examining developmental mechanisms of thinning and an opportunity to test the relationship between cortical thinning and morphology. After localizing functional regions of interest (fROIs) in each participant, we measured: (1) CT of each fROI, (2) T1 and MD in white matter adjacent to each fROI using qMRI and dMRI, respectively, (3) T1 and MD in gray matter of each fROI across intra-cortical depths, and (4) cortical curvature of each fROI. We compared these measurements across age groups to determine which factors develop and if so, whether these developments are region-specific or region-general. Finally, we tested if cortical thinning is linked to development of T1, MD, or cortical curvature.
Results
First, we verified that data quality was not lower in children than adults. Due to excessive motion (> 2 voxels) during fMRI, data from 3 out of 30 adults and 1 out of 27 children were removed from further analysis. In the remaining 26 children and 27 adults, we found no significant differences in motion during scan (F2,100=2.61, p>0.05, Fig. S1b). fMRI timeseries signal-to-noise ratio (tSNR) across VTC was higher in children than adults (main effect of age: F2,100=37.49, p<0.05, Fig. S1c). For dMRI, data from 2 children, 4 preteens, and 3 adults were excluded because of excessive motion above 2 voxels (1 voxel = 2 × 2 × 2 mm3), leaving 20 children (ages 5-9, N=9, ages 10-12, N=11) and 24 adults (ages 22-28) for dMRI analyses. After these exclusions, there was no significant difference across age groups. These quality control analyses demonstrate that (i) we can obtain high quality measurements in children and (ii) any developments that we report are likely not driven by nonspecific differences between age groups such as head motion or lower tSNR.
In each participant’s brain, we defined fROIs selective for faces (pFus-faces and mFus-faces, which together are referred to as the fusiform face area34,35), characters (pOTS-chars and mOTS-chars located in the occipital temporal sulcus (OTS), and also referred as visual word form area, VWFA, 1 and 2, respectively36, 37), and places (CoS-places, also referred as parahippocampal place area38) from the localizer experiment. We were able to localize these fROIs even in our youngest participants (Fig. 1a and Supplemental Table 1). Additionally, in each subject we generated whole brain maps of T1 and MD, and measured mean T1 and MD values in each subject’s fROIs and their adjacent white matter.
VTC thins from childhood to adulthood
Maps of CT on each subject’s cortical surface were generated using FreeSurfer39 from T1 images. Our measurements of CT are consistent with prior ex vivo40 and in vivo studies2 (Fig. S2). In general, CT decreased with age across VTC, except for a region near the temporal pole in which CT increased with age (interaction between age of subject and anatomical parcel, F16,900 = 19.48, p<0.05, Figs. S3a,b).
Measurements of CT in fROIs of VTC showed that cortex thinned from age 5 to adulthood in face-, character-, and place-selective regions (main effect of age, 3-way repeated measures analysis of variance (ANOVA) with factors: age of subject (5-9/10-12/22-28 years), fROI (face-/place-/character-selective), and hemisphere (left/right), F2,384 =78.74, p<0.05, Fig. 1b). Children’s cortex (ages 5-9) was on average 0.73±0.11 mm thicker than adults’. The smallest difference was observed in right CoS-places (0.46 mm) and the largest difference in right mFus-faces (0.97 mm). Additionally, CT varied across fROIs (main effect of fROI, F2,384 =21.00, p<0.05), whereby cortex was thinnest in CoS-places (2.56±0.45 mm), which lies close to the fundus of the collateral sulcus (CoS), and thickest in mFus-faces (2.97±0.65mm), which lies on the lateral fusiform gyrus. There was also a small but significant hemispheric difference, as left hemisphere fROIs were thinner than the right ones (F1,384 =4.99, p<0.05). As cortex thins, the volume of all fROIs did not decrease. There was no significant development in the volume of character- and place-selective regions (p=n.s.), and a developmental increase in the volume of face-selective regions (F2,206=5.2, p<0.05).
In functionally defined white matter (FDWM), T1 relaxation time and MD decrease from childhood to adulthood
As we found substantial development in CT, we first examined if these developments may be related to changes in white matter properties adjacent to each fROI. Therefore, we dilated each fROI 5 mm into the white matter (Fig. 2a and Methods), which we refer to as functionally-defined-white-matter (FDWM41). We estimated T1 (Fig.2b) and MD (Fig. 2c) in FDWM of face-, character-, and place-selective fROIs in each participant. Results were similar for fROIs from the same domain. Thus, in subsequent main figures we show a single fROI from each domain: face-selective mFus-faces, character-selective mOTS- chars, and place-selective CoS-places. Data from additional face (pFus-faces) and character fROIs (pOTS- chars) are in Figs. S4, S6.
Results show differential development of T1 in FDWM (interaction between age of subject and fROI, F4,384=2.54, p<0.05, Figs. 2d and S4a). T1 progressively decreased from age 5-9 to 10-12 to adulthood in the FDWM near face- and character-selective regions (Fs>18.12, ps<0.05). However, there was no significant change in T1 in FDWM of place-selective cortex after age 5 (F2,98=2.3, p=n.s.) In FDWM where T1 decreased, MTV concomitantly significantly increased with age. (Fig. S5).
MD in FDWM also revealed differential development (interaction between age of subject and fROI, F4,320=4.64, p<0.05). MD monotonically decreased in FDWM near face- and character-selective fROIs from age 5 to adulthood (Fs>8.77, ps<0.05, Figs. 2e and S4b). However, there was no development in MD of FDWM near CoS-places (F2,80=2.57, p>0.05). Decreases in both T1 and MD in FDWM next to face- and character-selective fROIs (but not place-selective cortex) are consistent with the hypothesis that development of white matter near these fROIs is associated with increased myelination.
In mid and deep cortical depths, T1 and MD decrease from childhood to adulthood
Next, we tested if there are tissue developments in cortex. We evaluated T1 and MD in each fROI across cortical depths in equidistant steps (10 steps for T1 and 8 steps for MD), extending from the pial surface to gray-white matter boundary, with the last two steps inside the white matter (Fig. S6).
Data reveal three findings. First, mean T1 decreased with age in face- and character-selective cortex (Fs>4.37, p<0.05), but not in place-selective cortex (F2,98=0.35, p=n.s.). Second, examination of T1 across cortical depth shows that T1 near the gray-white matter boundary was lower than T1 in superficial layers of gray matter (main effect of cortical depth in a 4-way ANOVA with factors of age of subject, fROI, hemisphere, and depth, see online methods, F4,1920 = 3450.47, p<0.05, Fig. 3a). Third, across cortical depths, in face- and character-selective fROIs, the largest development in T1 occurred away from the superficial pial surface and was more prominent in mid cortical depths (interaction between age of subject and depth in 4-way ANOVA with factors of age of subject, fROI, hemisphere, and depth, F8,1920 = 17.51, p<0.05, Fig. 3a). In contrast, in CoS-places, T1 curves largely overlapped across age groups. Data from pFus-faces and pOTS-chars are in Fig. S7a.
Notably, T1 curves were shifted leftward in adults compared to children. In other words, comparable T1 values were more superficial in adults than children. For example, T1 at the gray-white boundary in 5–9 year-olds was equivalent to T1 at ∼80% depth in adults. To quantify this shift, we fitted a maximum log-likelihood function to each subject’s T1 curve, and estimated the inflection point of the curve, which reflects the depth at which slope is maximal (Fig. 3b). Results revealed that the depth of this inflection point was closer to the pial surface in adults than children, with larger development in face- and character-selective fROIs than in place-selective fROIs (significant interaction between age of subject and fROI, F4,384 = 3.89, p<0.05).
The development of MD across intra-cortical depths is more complex as gray matter is not as directionally organized as white matter (Figs. S7b, S8a). Across age groups, MD decreased from the pial surface to the gray-white matter boundary (interaction between age of subject and cortical depth: F6,1248=8.26, p<0.05). Analysis of MD revealed differential development across fROIs (significant interaction between age of subject and fROI, F4,1248=4.01, p<0.05). There were prominent developmental decreases in MD in face- and character-selective fROIs (Fs>4.13, p<0.05), but not in place-selective cortex (F2,324=0.39, p=n.s.). Additionally, within face-selective fROIs, MD development was observed in deeper cortical layers than superficial layers (F6, 540=4.82, p<0.05).
Together, results show that both T1 and MD in cortex decrease from age 5 to adulthood suggesting microstructural tissue growth and not tissue loss.
Development of T1 and MD in gray matter correlates with CT
Does tissue growth in face- and character-selective regions correlate with apparent cortical thinning? We reasoned that if thinning relates to developments in tissue properties then there will be a significant positive correlation between CT and T1 (or MD). To test this, we quantified in each fROI the relationship between CT and T1 (or MD) across cortical depths. In face-selective (bilateral pFus-faces and mfus-faces) and character-selective fROIs (bilateral pOTS-chars and left mOTS-chars), we found a significant positive correlation between CT and T1 (Rs>0.48, ps<0.05) and between CT and MD (Rs>0.36, ps<0.05). The highest correlation was between CT and T1 at ∼50% cortical depth from the pial surface (Fig. 4) and between CT and MD at ∼80% cortical depth (Fig. S8b). The correlation between CT with T1 remained significant after age was partialled out (Fig. 4; same for MD, Fig. S8b). In CoS-places, where we did not find development of tissue properties, we also did not find a significant correlation between CT and T1 (right: CoS-places, R=0.3, p=n.s.) or CT and MD (Rs<-.06; ps=n.s.), except for a significant correlation between CT and T1 at ˜50% depth in left CoS-places (R=0.47, p<0.05; significant after age partialled out). Together, these analyses suggest that in mid and deep cortical depths of face- and character-selective fROIs, participants with more tissue (lower T1 and MD) have a thinner cortex than those with less tissue (higher T1 and MD).
Cortical curvature and surface area are linked to developmental thinning in CoS-places
Our measurements did not find tissue development in CoS-places even as the CoS-places thins from age 5 to adulthood (Fig. 1b). Thus, we next asked if cortical thinning in CoS-places is linked to morphological changes including changes in cortical curvature and surface area.
As CoS-places is in a sulcus33 and sulcal fundi are typically thinner than sulcal walls and gyri39, we examined if developmental changes to the curvature of CoS-places or developmental shifts in the location of CoS-places towards the fundus may explain cortical thinning. In morphological terms, the bottom of the sulcus, or the fundus, has the largest positive curvature (more concavity) than the sulcal wall or a gyrus (less concavity). Analysis of cortical curvature implemented in FreeSurfer revealed that adults’ CoS-places was on a more concave surface than children’s (main effect of age with factors age of subject and hemisphere, F2,98=3.47, p<0.05, Fig. 5a). In contrast, there was no curvature development in face- or character-selective fROIs (Fs<1.45, ps=n.s.). Further, we found a significant negative correlation between CT and curvature in CoS-places (Fig. 5b, right: R=-0.43, p<0.05, significant when age partialled out; left: R=-0.34, p<0.05, n.s. when age partialled out). That is, more concave CoS-places was associated with thinner cortex. There was no significant correlation between CT and curvature in bilateral pFus-faces, mFus-faces, and left pOTS-chars (Rs>-0.3, ps>.05), but we found a negative correlation in left mOTS-chars and right pOTS-chars (Rs<-0.33, ps<0.05; significant when age partialled out). Overall, results suggest that adults’ thinner CoS-places might be situated in the fundus of the CoS as compared to children’s thicker CoS-places.
To explore this prediction, we visualized the location of CoS-places on the cortical sheet, by generating a group map of CoS-places in children (N=26) and adults (N=27) using cortex-based alignment to the FreeSurfer average brain. As shown in Fig. 5c for the right hemisphere, while some of the children’s CoS-places is within the fundus of the CoS, there is variability across subjects along the lateral-medial axis of the CoS. In comparison, adults’ CoS-places is less variable and is consistently localized in the CoS fundus.
Change in structural-functional coupling across development is not associated with growth in the volume of CoS-places (p=n.s.). Thus, we tested for morphological changes in the CoS. That is, we examined (i) if cortex stretches during development42, and (ii) if areal expansion is coupled with thinner cortex. We derived the CoS from the FreeSurfer anatomical parcellation in each subject, and evaluated if its surface area, including and surrounding CoS-places, develops. Surface area of the CoS expands with age (main effect of age, F2,100=4.07, p<0.05, Fig. 5d) and it was negatively correlated with CT in both hemispheres (Rs<-.41, ps<0.05; significant when age partialled out). Together, results suggest that thinning in place-selective cortex is linked to morphological developments whereby the cortex stretches and leads to the location of adult’s CoS-places to be consistently in the fundus.
Increased myelination as likely mechanism underlying cortical thinning in face- and character-selective areas
What cellular mechanism underlies tissue growth in face- and character-selective fROIs? We hypothesized that one tissue compartment that may affect cortical development of T1 and MD may be myelin. However, little is presently known about the myelination of axons in VTC, and how deep into cortex they penetrate. Addressing this question requires measurements of myelin in histological tissue slices of postmortem brains.
While we were unable to obtain pediatric postmortem tissue slices to measure myelin development, we were able to leverage tissue differences across fROIs as a proxy of developmental myelination changes and to examine the effect of myelin on in vivo cortical tissue measurement. That is, another aspect of our results revealed that in adults, but not children, relaxation rate (R1=1/T1) was higher in face-selective fROIs (bilateral pFus-faces and left mFus-faces) than in CoS-places (Fig. 6a). In other words, tissue of face- and place-selective regions is undifferentiated in childhood, and development leads to differentiated R1 across adult face- and place-selective fROIs. We reasoned that if myelin contributes to the development of R1 in face-selective cortex, then face-selective cortex in adults should be more myelinated than place-selective cortex, and especially so in deeper cortical layers.
To identify these regions in postmortem histological slices, we need anatomical markers. Prior research from our lab identified reliable anatomical landmarks that predict face-selective and place- selective regions in VTC: the mid-fusiform sulcus (MFS) predicts face-selective regions32 and the intersection of the anterior lingual sulcus and CoS predict CoS-places33 (Fig. 6b). We identified these landmarks in coronal sections (30 microns thick) from 5 postmortem brains that contained the fusiform gyrus (FG), MFS, CoS, and calcarine sulcus (Fig. 6c), and then stained histological sections containing these landmarks for myelin using modified Gallyas stain43 (Figs. 6d-f). We estimated myelin as optical density (OD), which indicates the index of light refracted through the tissue (more myelin produces a darker stain, Fig. 6d, Fig. S9).
Myelin stained coronal sections (example: Fig. 6d; all brains: Fig. S9) revealed striking laminar differences across cortex. Within cortex, deeper cortical layers had higher myelin than superficial layers. The differentiated pattern is also evident in high-resolution (50 microns) T2* weighted image of the same brain (Fig. 6g). Zooming in on the CoS (Fig. 6e) and FG (Fig. 6f), we noted two interesting observations: (1) myelinated fibers project roughly orthogonal to the cortical surface and the depth of myelin penetration varies between FG and CoS. Myelin penetrates about ∼54% into the cortex in FG, but only penetrates ∼36% in CoS. This is consistent with prior findings using equivolume models44, 45 that report deeper myelin penetration in cortical layers of gyri than sulci to compensate for thickness. (2) The amount of myelin is greater (darker stains) in FG than in CoS and this difference is not just due to sulcal vs. gyral differences in myelination, as the Stria of Gennari in V1 (Fig. 6d) in calcarine sulcus, is strongly myelinated.
We quantified these observations by measuring normalized optical density profiles along equidistance depths from the pial surface to white matter for MFS/FG and CoS in each post mortem brain. Myelin content increased from superficial layers to deeper cortical layers (Fig. 6h). Importantly, estimated myelin in MFS/FG was greater than in CoS in depths away from the pial surface in 4 out of 5 brains (Fig. 6h), which is largely consistent with our in vivo R1 data (Fig. 6a).
Thus, for the first time, using histological analyses of the MFS/FG and CoS, we show that myelination is a likely mechanism underlying differences in adult in vivo T1 measurements across these regions. Although we do not have pediatric data, results are consistent with the idea that higher myelination in face- than place-selective cortex in adults is due to increased myelination across development.
Discussion
Combining CT, qMRI, dMRI, and histology measures for the first time, our study offers the first mechanistic understanding of development of cortical thickness, which is one of the most widely used measures of brain development. First, we found no evidence of pruning after age 5 in any region in VTC. Moreover, in face- and character-selective regions, qMRI and dMRI revealed decreases in T1 and MD, both of which occurred in mid and deep gray matter and in the adjacent white matter. These decreases are associated with more tissue, rather than less tissue as predicted by pruning. Second, myelin staining in postmortem brains showed that myelin penetrates into gray matter particularly in the FG, providing evidence that myelin is a significant contributor to T1 and MD development in face- and character- selective regions, measured in the present study. Third, we found heterogeneous mechanisms of cortical thinning across VTC. Different than face- and character-selective regions, there was no development of gray or white matter of the CoS-places, instead thinning was coupled with morphological changes. Critically, our results suggest a rethinking of mechanisms of brain development, providing evidence that after age 5, myelin increases in the FG and OTS, and the CoS undergoes morphological changes.
Several innovations in our study have enabled the advancement of understanding mechanisms of cortical thinning across development. First, the use of individual subject analyses with no spatial smoothing and tissue measurements across cortical depths has increased the precision of measurements from centimeters, in prior work employing group analyses, smoothing, and large macroanatomical landmarks1-5, to millimeters. Second, leveraging our understanding of functional-structural relationships in VTC32, 46, we identified relevant landmarks in postmortem brain slices and validated in vivo measurements with ground truth measurements of myelin. Third, we obtained independent measurements of T1 from qMRI18, 19, 47 and MD from dMRI21 in both gray and white matter. These complementary measurements yielded consistent results of growth in both tissue types, strengthening our findings. These advancements not only provide striking empirical evidence supporting developmental theories of increased myelination in deeper gray matter and its adjacent white matter2,11, 13, but also underscore the utility of obtaining multiple measurements in the same individual to glean insights into developmental mechanisms.
Our data provide evidence that increased myelination of axons during childhood is a key source of cortical thinning in VTC after age 5. While prior imaging studies documented pervasive thinning throughout childhood, which we replicate (Figs. 1,S2,S3), till now the underlying mechanisms have been hypothetical. Some researchers have speculated that cortex thins due to synaptic pruning and cell loss3. However, others argued that thinning is likely not due to tissue loss because they found developmental increases in gray and white matter volume1, 2. Therefore, they speculated that thinning may be due to a shift in the gray-white boundary from increased myelin. Three of our findings support the latter hypothesis for the development of face- and character-selective regions. First, we found age-related decreases in T1 and MD in white matter next to these regions. As myelin explains 90% of the variance in T1 of white matter48, this finding is likely explained by increased myelination. Second, in cortex, we found decreases in T1 and MD far from the pial surface. Third, myelin staining in ex vivo adult brains shows that myelin penetrates deep into cortex, especially in the FG. Thus, our findings provide evidence for myelin proliferation into the cortical neuropil during childhood rather than tissue loss, which suggests a rethinking of the term ‘cortical thinning’ to describe structural development of gray matter. In other words, our data suggests that increased myelin during development brightens the gray-white matter interface, shifting the boundary deeper into cortex in adults. While our data suggest that myelin is a key contributor, our prior measurements and simulations indicate that tissue growth affecting T1 cannot be exclusively due to myelin increases26. Growth in additional microstructures including glia22, 49, dendritic arbors, synapses50, and iron development in glia and myelin20 likely contribute to T1 development.
Is it possible that pruning still occurs during development? One possibility is that pruning occurs earlier in development, e.g. during infancy7, 8. A second possibility is that pruning after age 5 is smaller in magnitude compared to myelination, and therefore the net effect on MRI measurements is dominated by myelination. Third, it is possible that our voxels and field strength do not have sufficient sensitivity to measure pruning-related developments where the cortex is thin, highly gyrified, or strongly myelinated20. Future measurements with high-field MRI and submillimeter resolution, and histology of pediatric and adult postmortem brains can test these possibilities51.
Another interesting finding is that different mechanisms underlie apparent thinning in face- and character-selective cortex vs. place-selective cortex, which is merely 2 cm away. In CoS-places, there was no development in either gray or white matter tissue. Instead, CT was correlated with the curvature of CoS-places and the surface area of CoS. This suggests that the CoS may stretch, perhaps deepening the sulcus in adulthood, resulting in cortical thinning. Several mechanistic forces may alter this cortical landscape during childhood, including axonal tension15, cytoarchitectural patterning52, and differential elasticity properties between white and gray matter16. Our findings that morphological changes play a role in cortical thinning, highlight that mechanical forces, an overlooked factor of brain development, should be considered not only during embryonic development, but also during childhood.
What may explain the differential mechanisms of cortical tissue development across VTC? One possibility is that different mechanisms affect development of sulci vs. gyri. However, our data suggest that this is not the case. Both CoS-places and mOTS-chars are in sulci and the latter exhibits tissue growth, while the former does not. Likewise, both CoS-places and V1 are in sulci and the latter has significant myelination (Stria of Gennari in V1, Fig. 6d), while the former does not. Another possibility is interplay between functional and structural development. Face-26,28-30, 53,54 and character-selective regions31,54-56 show a protracted development compared to object- and place-selective regions in VTC. Research on oligodendrocytes and their progenitor cells indicate that development of myelin is activity-dependent57. Based on this work and our findings, we propose a new intriguing hypothesis: cortical regions that exhibit protracted functional development will exhibit protracted myelin development, and consequently prolonged apparent cortical thinning. This hypothesis can be tested in future research.
In conclusion, a major goal of neuroscience is to understand mechanisms of brain development. Our study demonstrates the feasibility of evaluating tissue properties in gray and white matter in children and adults using 3T MRI. Additionally, our study is transformative because it underscores the significance of measuring microstructural and morphological changes to brain tissue across the lifespan, rather than just volumetric and thickness measures. Further, as increased myelin during development brightens cortex, our measurements have implications for (i) widespread algorithms estimating CT, brain volume, and tissue type (gray vs. white matter) from MRI, (ii) generating age- appropriate brain atlases, and (iii) large-scale studies of brain development and neuroanatomy including the Adolescent Brain Cognitive Development (ABCD)58 study and the Human Connectome Project (HCP)59, as these studies will impact future policies that promote the health and well-being of children. Finally, our data have important ramifications for understanding typical60 and atypical brain development as well as for clinical conditions implicating myelin and morphology including autism61, dyslexia62, and multiple sclerosis63.
Authors’ Contributions
VSN designed fMRI experiment, collected and analyzed fMRI, qMRI, dMRI, analyzed post mortem data, and wrote the manuscript; JG contributed to experiment design, MR data collection, and analysis; MB and BJ contributed to the collection and analysis of MR data. ZZ contributed to the analyses of dMRI. SC contributed to defining anatomical ROIs. KSW contributed to discussions on post-mortem data. EK, CJ, and NW contributed to the collection of post mortem brains, MR of post mortem brain, and generation of histological myelin stains. KGS oversaw all components of the study, and wrote the manuscript. All co-authors have read and approved the submitted manuscript.
Competing interests statement
Authors have no competing interests.
Online Methods
Participants
27 children (14 females, ages 5-12) and 30 adults (11 females, ages 22-28) participated in our study. Children were recruited from the Palo Alto school district, through flyers and online advertisements. All children attended public schools at the time of recruitment. Adult subjects are Stanford University affiliates. MRI data was collected using a 3T scanner in the Center for Cognitive and Neurobiological Imaging (CNI) at Stanford University. All subjects had normal or corrected-to-normal vision and provided written, informed consent. Protocols were approved by the Stanford Internal Review Board on Human Subjects.
General training and scanning procedures
Multiple sessions
Adults and children participated in several scanning sessions on different days, to obtain quantitative MRI (qMRI), diffusion MRI (dMRI), and functional MRI (fMRI) data.
Training
Children underwent a separate training session, prior to scanning, inside an MRI scanner simulator. Children were trained to remain still by monitoring their motion, using a motion sensor attached to their forehead and providing them with feedback regarding the amount of head movement they made. Children who stayed still during the mock session were invited back for MRI scans.
Acquisition
In vivo anatomical and functional data were acquired in each subject using a 3T GE Signa MRI scanner, using a custom-built phase array 32-channel, head coil using methods as in prior studies18,26. During MRI scanning, participants lay supine inside the magnet. Visual stimuli were projected onto a monitor and were viewed through an angled mirror mounted above the participant’s head.
qMRI data
Acquisition
Subjects were scanned using spin-echo inversion recovery with an echo-planar imaging (EPI), read-out (SEIR-EPI). This scan was done with a slab-inversion pulse and spatial-spectral fat suppression. We used 2mm2 in-plane resolution with a slice thickness of 4 mm and the EPI readout was performed using 2X acceleration. QMRI parameters were measured from spoiled-gradient echo images acquired with different flip angles (α = 4°, 10°, 20° and 30°, TR = 14 ms, TE = 2.4 ms) and a voxel resolution of 0.8×0.8×1 mm3, which was resampled to 1 mm3 isotropic. For SEIR-EPI, the TR was 3 s. The echo time was set to minimum full; inversion times were 50, 400, 1200 and 2400 ms. Anatomical data were aligned to the AC-PC plane. From the qMRI, we generated whole-brain T1-weighted anatomy18, 26.
Quality assurance
Subjects whose anatomical data appeared blurry or showed ringing due to motion were asked to return for a second scan on a different day (N subjects = 4, mean age = 6.75 years). The second scan was of sufficient quality and was used for subsequent visualization and cortical surface reconstruction.
General preprocessing and analyses
The spoiled-GE and the SEIR scans were processed using the mrQ software package in MATLAB to produce the T1 and MTV maps. The mrQ analysis pipeline corrects for RF coil bias using SEIR-EPI scans, producing accurate proton density (PD) and T1 fits across the brain. MTV maps were produced by calculating the fraction of a voxel that is not MR visible -water (CSF voxels are taken to be nearly 100% water). The full analysis pipeline can be found at (https://github.com/mezera/mrQ)18.
Reconstruction of cortical surfaces
T1 images underwent automated cortical surface reconstruction using FreeSurfer (http://surfer.nmr.mgh.harvard.edu/). The anatomical images were segmented into white and gray matter. White matter surfaces were inspected and manually fixed for missing or mislabeled white matter voxels using ITK-SNAP (http://www.itksnap.org/). A mesh of each participant’s cortical surface was generated from the boundary of the white and gray matter. This mesh was inflated for visualization of activations inside the sulci.
Generation of cortical thickness, curvature maps, and surface area
For each subject, we used FreeSurfer’s automated algorithm to define the gray-white matter boundary, pial surface, and cortical thickness (CT) maps. CT (in mm) is measured as the distance between the gray-white boundary and pial surface39. FreeSurfer’s algorithm also generates curvature files, which contain information about the amount of curvature at each vertex (i.e., sharper the curve, higher the value) and information whether the vertex is on a sulcus or gyrus (indicated by a sign: a positive sign indicates a sulcus, or concave fold, while a negative sign indicates a gyrus, or convex fold). Mean curvature is made up of the average of principal curvatures derived from the inverse of the radius of the osculating circles at each point on the surface on the gray-white matter junction. Surface area in mm2 of the collateral sulcus was calculated from FreeSurfer’s anatomical parcellation of the collateral sulcus for each participant and both hemispheres using FreeSurfer’s mris_anatomical_stats and aparcstats2table functions.
Analysis of CT in FreeSurfer’s anatomical parcellations
To replicate previous results of cortical thinning in large anatomical expanses of the brain1, 2, we measured CT in nine anatomical parcellations of the left and right VTC obtained from FreeSurfer’s algorithm for each subject (see Fig. S3b). These include: occipital pole, calcarine sulcus, inferior temporal gyrus, occipital temporal sulcus, inferior temporal sulcus and gyrus, collateral sulcus, fusiform gyrus, parahippocampal gyrus, and temporal pole. We generated an “average brain” from our 27 adults and 26 children using FreeSurfer’s make_average_subject algorithm. Next, each subject’s brain and their corresponding CT maps were transformed into this average brain space using FreeSurfer’s mris_preproc and mri_surf2surf algorithms. We then measured average CT per anatomical region, across all subjects in an age group, and conducted a 3-way ANOVA with age of subject, hemisphere, and parcellation as factors.
dMRI data
Acquisition
Consistent with methods in our prior study41, in each subject, we obtained two whole-brain, diffusion-weighted, dual-spin echo-sequence (60 slices, TE = 96.8 ms, TR = 8000 ms, 96 diffusion directions, b0 = 2000s/mm2, voxel size 2×2×2 mm3). Ten non-diffusion-weighted images were collected at the beginning of each scan.
Quality assurance
We assessed motion during dMRI from tensor images. Two 5-9 year olds, four 10-12 year olds, and three 22-28 year olds moved during dMRI excessively (motion > 2 voxels, 1 voxel = 2 × 2 × 2 mm3). Data from these participants was excluded from subsequent dMRI analyses. MD data is consequently reported for 20 children (ages 5-9, N= 9, ages 10-12, N=11) and 24 adults (ages 22-28).
General preprocessing and analyses
Diffusion data analyses were done using mrDiffusion (http://white.stanford.edu/software). Non-diffusion-weighted images were averaged, and diffusion-weighted (DW)images were registered to this mean image using a two-stage model with a coarse-to-fine approach that maximized mutual information. DWIs were motion and eddy-current corrected and aligned to each subject’s whole brain anatomy using SPM8 (http://www.fil.ion.ucl.ac.uk/spm/software/spm8/). Tensors were fit to each voxel using a least-squares algorithm that removes outliers. For each subject, we obtained mean diffusivity (MD) maps from the tensors files.
fMRI data
Acquisition
Data were acquired using a T2*-sensitive gradient echo spiral pulse sequence with a resolution of 2.4 × 2.4 × 2.4 mm3, TR = 1000 ms, TE = 30 ms, flip angle = 76°, and FOV = 192 mm. We collected 48 oblique slices, oriented parallel to the superior temporal sulcus, using a multiplexing technique allowing whole-brain coverage of functional data. The same prescription was used to obtain anatomical T1-weighted images, which were used to align functional data with the whole brain high-resolution anatomical volume of each participant.
Functional localizer
In order to localize face-, character-, and place-selective regions of interest (ROIs), subjects participated in a localizer experiment based on our prior methods26, 37, 53. Subjects participated in 3 runs of a functional localizer experiment (5.24 min/run) with 78, 4s blocks in each run. Subjects viewed gray-scale stimuli which were blocked by category. Images consisted of two subtypes from each of five categories: characters (numbers and pseudo-words), bodies (limbs and headless bodies), human faces (child faces and adult faces), places (houses and indoor scenes), and objects (guitars and cars) (see Fig. S1a for example images per category). Each image was shown only once during the experiment. In each 4 s block, different stimuli from one of the above categories were shown at a rate of 2 images per second. Blocks were counterbalanced across categories and also with baseline blocks consisting of a blank, gray background. During the scan, subjects fixated on a central dot and pressed a button when a phase-scrambled, “oddball” image appeared randomly in a block (˜33% of the blocks).
Quality assurance
(1) Motion. Data of each subject were corrected for within-run and between-run motion. Only runs with motion of less than two voxels were included in the study (as in our prior studies26,53). To test for development, we performed a 2-way ANOVA with factors, age of subject and motion type. (2) Time series signal to noise ratio (tSNR). As a control to test if developmental changes were driven by smaller signal to noise ratio in children than adults, we measured tSNR across children’s and adults’ left and right VTC ROIs. For each subject, we defined left/right VTC as the boundary between the posterior transverse collateral sulcus (ptCoS) to the anterior tip of the MFS, including the extreme edges of the CoS and OTS. tSNR was computed for each voxel as follows: tSNR = mean(time series)/standard deviation (time series) and then averaged across all voxels of the VTC, across runs and across subjects in an age group. We performed a 2-way ANOVA with factors, age of subject and hemisphere to test development in tSNR.
General processing and analyses
Localizer data were analyzed using code written in a MATLAB based mrVista toolbox (http://github.com/vistalab) as in our prior publications26, 37, 53, 54. Data were not spatially smoothed and were analyzed in each subject’s native brain space. The time courses of each voxel were converted from arbitrary scanner units into units of % signal change.
General Linear Model (GLM)
For each subject’s data, we ran a GLM to model each voxel’s time course. The experimental design matrix was convolved with the SPM hemodynamic response function (HRF) (http://www.fil.ion.ucl.ac.uk/spm) to generate predictors. Using a GLM to fit the predictors to the data, we estimated the response amplitudes for each condition (betas) and residual variance of each voxel’s time course. We used beta values and residual variance from the GLM to generate contrast maps comparing responses in different conditions.
Definition of functional regions of interest (fROIs)
Functionally selective voxels were defined as voxels that responded more to images of a selected category than to images of other categories (t>3, voxel level) during the localizer scan. Thus, in each subject, we defined the following regions in the ventral stream. Spatially contiguous clusters of face-selective voxels that responded more strongly to faces (child faces and adult faces) than other stimuli and were located in the posterior lateral fusiform gyrus were defined as pFus-faces. Face-selective voxels near or overlapping the anterior tip of the mid-fusiform sulcus (MFS) were defined as mFus-faces (pFus- and mFus-faces are collectively referred to as the fusiform face area, FFA34). Clusters of voxels that responded more strongly to places (house and corridors) than other stimuli and were located in the collateral sulcus (CoS) were defined as place-selective CoS-places also referred to as the parahippocampal place area (PPA)38. Clusters of voxels that responded more strongly to characters (numbers and pseudo-words) than other stimuli and were located in the posterior OTS were defined as pOTS-characters, also referred to as visual word form area 1{Stigliani, 2015). More anterior character-selective voxels in the occipital temporal sulcus were defined as mOTS-characters also referred to as visual word form area 237 (mOTS- and pOTS-words are collectively referred to as the visual word form area, VWFA36).
Behavior
We collected behavioral data during the scan in an odd-ball task, in which participants were asked to press a button when they viewed a scrambled image on the screen, which appeared in 33% of the trials. Due to equipment malfunction, however, we were able to collect data from 10 children (ages 5-12, 2 females) and 17 adults (ages 22-26, 8 females). Performance accuracy in the oddball task during fMRI was not significantly different (t(24)=0.42, p=n.s., two-tailed) across children (78.0±7.78%) and adults (81.43±4.51%).
In vivo analyses examining anatomical features of functional regions
Analysis of CT in functional ROIs of VTC
Data were analyzed using a MATLAB-based, mrVista toolbox. In brief, each subject’s CT surface maps from FreeSurfer’s algorithm were transformed to volume nifti files using FreeSurfer’s mri_surf2vol algorithm. For each subject, we averaged CT per fROI and hemisphere, across all subjects within an age group. We performed a 3-way ANOVA with factors, age of subject, fROIs, and hemisphere.
Analysis of T1, MTV, and MD in functionally-defined white matter (FDWM)
Each individual subject’s fROI was dilated into its neighboring white and gray matter by 5mm, to obtain T1, MTV, and MD measures of its functionally-defined white matter. Average T1, MTV, and MD were calculated for these FDWM voxels, and averaged across subjects within an age group. We performed a 3-way ANOVA with age of subject, hemisphere, and fROIs as factors per measure.
Analysis of T1 /MD by intra-cortical depth
Using FreeSurfer’s mri_surf2vol algorithm, each subject’s FreeSurfer-based white matter surface was converted to volume niftis. The volume was projected into gray matter at increasing 20% steps, starting from two locations below the gray-white boundary (to include some white matter just below the boundary) to two locations above the pial surface (to include the entire pial surface). As thickness of the cortical ribbon varies between 2mm to 4mm in the VTC, each increasing depth is between 0.4 mm – 0.8 mm from the gray-white surface normal. For each subject’s dilated fROI, we averaged T1 values of those voxels that intersected with each cortical layer, and obtained T1 values per cortical depth. We performed 4-way ANOVAs with age of subject, fROI, hemisphere, and cortical depth as factors. Similar analysis was conducted for MD, with the exception that the cortex was divided into 8 steps rather than 10 steps as MD maps were 2mm voxel resolution. For the statistical analysis of variance using depth as a factor, we used every other step instead of all steps to ensure that they reflect data from non-overlapping voxels (5 out of 10 steps for T1 and 4 out of 8 steps for MD). For each fROI, we correlated each subjects’ CT with T1 (or MD) at each cortical depth (Pearson’s correlation coefficient).
Correlation between anatomical curvature and CT
Each subject’s curvature maps from FreeSurfer’s autosegmentation algorithm were transformed into volume nifti files using the FreeSurfer-based mri_surf2vol algorithm. Data were analyzed using MATLAB-based mrVista toolbox. Per participant, we calculated average curvature in each fROI and examined if there was a significant correlation (using Pearson’s correlation coefficient) between CT and curvature per fROI.
Post mortem tissue data
Histology (post mortem tissue blocks)
Blocks of five post mortem human brains were provided by the Brain Banking Centre Leipzig of the German Brain-Net (GZ 01GI9999-01GI0299), operated by the Paul Flechsig Institute of Brain Research (University of Leipzig). Information on sex, age of death, cause of death and post mortem interval before fixation (PMI) is provided in Supplementary Table 2. The entire procedure of 1) case recruitment, 2) acquisition of the patient’s personal data, including protocols and informed consent forms, 4) the performance of the autopsy, and 5) the handling of the autopsy material have been approved by the responsible authorities. Following the standard Brain Bank procedures, tissue blocks from occipital cortex were extracted, immersion-fixed in 4% paraformaldehyde in phosphate buffered saline (PBS) pH 7.4 for at least 6 weeks. Prior to MR scanning, blocks were washed in PBS with 0.1% sodium azide for at least 24 h in order to avoid biasing the effects of paraformaldehyde on MR relaxation rate. All tissue blocks contained a part of the calcarine fissure, collateral sulcus, and fusiform gyrus.
MR Scanning
MR scanning of post mortem tissue samples was performed on a 7T Siemens scanner (Magnetom, Siemens, Erlangen, Germany) equipped with a custom made 2-channel transmit-receive coil with an inner diameter of 6 cm. Samples were placed in a spherical container filled with Fomblin (Sigma Aldrich).
Tissue sectioning
For histological processing, tissue blocks were cryoprotected in 30% sucrose/PBS with 0.1% sodium azide. Sections of 30 microns thickness were made on a cryomicrotome (Leica SM2000R with freezing unit Zeiss KS34) and collected in PBS with 0.1% sodium azide.
Semi-quantitative myelin mapping in histological sections
For myelin staining, every 20th section was taken out of the series (in total: 6 to 9 sections per case), mounted on glass slides, and air dried. Sections were rehydrated in distilled water and stained following the classical protocol of the Gallyas method43 with slight adaptions65. In order to assure the complete visualisation of all myelin fibers, including those that are very thin, the developing step was extended to 40 min. After the procedure, slides were thoroughly washed in distilled water, dehydrated and cover slipped with Entellan/Toluene (Merck). Imaging was performed with a Zeiss Axio Scan, Z1 slide scanner with a 10x objective. All images were obtained under equal conditions to ensure comparable data for the optometric processing. Obtained RGB maps were logarithm transformed using the Lambert-Beer transformation and averaged across three color channels with corresponding weights to calculate optical density (OD) of the sections (i.e., OD =0.8*R + 1.2*G+ 1.5*B, where R, G, B are the red, green, and blue channels in a .tiff image file). Thus, OD is proportional to the density of staining chromophore and therefore, reflects local myelin density in the tissue.
Adult histology data vs. adult in-vivo T1 data in face- and place-selective cortex
General processing and analyses
Our strategy was to compare adult histology data to adult in vivo T1 data in face- and place-selective cortex. Thus, for each individual stained slice, we divided the cortex into equidistant depths from the pial surface to white matter. We achieved this by drawing a mid-line in MFS/FG and CoS (along its thickness) and projecting equidistant depths along the mid-line normals in two steps: 1) from mid-line to pial-surface and 2) mid-line to white matter. We then evaluated average OD at each depth (OD at each point was normalized with respect to OD in white matter), which in turn, reflects the measure of the amount of myelin per region.
Data availability
All code relevant to data analysis as well as the source data for the main findings (Figs. 1-6) will be made available upon request. Majority of the code used in this study was derived from functions available through the open-source code library: https://github.com/vistalab/vistasoft.
Footnotes
Conflict of Interest: The authors declare no competing financial interests.