Network-level structural covariance in the developing brain

Primary Sensory and Motor Networks.

Seeds within primary visual, auditory, and motor cortices produced SCNs with a distinctive pattern of age-related change (Fig. 1A). The primary visual cortex (right calcarine sulcus) anchored covariance maps that included medial and lateral occipital cortices in groups 1 and 2 and expanded to include inferotemporal, dorsal parietal, and lateral frontal regions in group 3, suggesting emergent covariance with downstream modality selective (ventral and dorsal visual streams) and heteromodal association systems (36). Group 4 showed a similar but much less distributed pattern compared with group 3, suggesting that structural covariance arising in adolescence may be pruned during progression to adulthood. Primary auditory cortex (Heschl's gyrus) covaried with contralateral Heschl's gyrus in Groups 1 and 2 and underwent significant expansion in group 3 to include bilateral supramarginal, inferior frontal, and superior parietal lobule regions as well as the precuneus. Group 4 again showed persistence of this overall pattern with notable topographic restriction compared with group 3. Primary motor cortex (right precentral gyrus) correlated with left precentral gyrus, bilateral postcentral gyri, and medial supplementary motor cortex in group 1 and persisted, with varying extent and significance, in groups 2–4. Group 4 again showed a relative decrease in network extent compared with group 3, although substantial covariance along the more rostral frontal convexity was first evidenced at this stage. In summary, we observed a common network trajectory pattern across the primary visual, auditory, and sensorimotor systems (Fig. 1B). The findings suggest that these networks are provisionally established by early childhood but undergo significant expansion in early adolescence before contraction or pruning in late adolescence.

Language-Related Speech and Semantic Networks.

Across the four age groups, language-related networks showed progressive increases in regional extent and distribution, consistent with the view that speech and semantic functions continue to mature throughout childhood (Fig. 2A). The left inferior frontal gyrus, pars opercularis (IFGo; containing Broca's area) seed anchored bilateral IFGo and ipsilateral dorsal anterior insula covariance in group 1, which expanded to include bilateral dorsal anterior insulae in group 2, expanded further to include dorsolateral prefrontal cortex in group 3, and expanded further still to include more extensive frontal regions, posterior superior temporal gyrus (Wernicke's area), and fronto-parietal sensorimotor and cingulate motor areas in group 4. A left temporal pole seed, chosen to map a network related to semantic knowledge (37, 38), covaried in group 1 with ipsilateral amygdala, contralateral temporal pole, and bilateral inferior temporal cortex. This pattern expanded to include bilateral insular regions in group 2 and contralateral amygdala, anterior cingulate, medial and lateral orbitofrontal, and frontopolar regions in group 3. Group 4 showed extensive inferior, mesial, and superior temporal and bilateral amygdalar, insular, orbitofrontal, and cingulate covariance, reflecting a previously identified adult ICN/SCN (14). Together, these findings suggest that language-related speech and semantic networks undergo continued expansion throughout childhood, with increasingly long-range covariance in early adolescence that progresses throughout the teenage years. The language-related patterns were reflected quantitatively by similar stage-wise increases in the number of significantly correlated voxels (Fig. 2B).

Salience, Executive Control, and Default-Mode Networks.

To assess networks associated with nonlinguistic social–emotional and cognitive functions, we measured structural covariance within an anterior cingulate-frontoinsular (FI) network related to salience processing (8), a dorsolateral fronto-parietal network critical for executive control and working memory (8), and a posterior cingulo-temporo-parietal DMN related to episodic memory, visual imagery, and mentalizing (7). Strikingly, the salience network and executive control network showed limited covariance patterns throughout initial stages of childhood, with far more distributed structural covariance in late adolescence (Fig. 3A). Specifically, the right frontoinsula (a major node within the salience network) anchored covariance maps that included bilateral frontoinsular and orbitofrontal regions in group 1 and bilateral insular, anterior cingulate cortex, and frontopolar regions in group 2. This SCN underwent significant extension in group 3 to include most of the medial frontal wall and bilateral insular cortex. In group 4, a robust and nearly contiguous SCN emerged comprised of extensive lateral and medial frontal, anterior cingulate, and anterior temporal cortices as well as bilateral amygdala. A seed in right dorsolateral prefrontal cortex (DLPFC; a major node within the executive control network) produced arguably the latest-maturing SCN, with only seed autocorrelation in groups 1 and 2 and contralateral DLPFC covariance in group 3. A much more extensive bifrontal covariance emerged in group 4, however, extending to dorsal anterior cingulate cortex, ventrolateral prefrontal cortex, and premotor areas. Right angular gyrus, a seed used to derive the DMN (14), showed limited covariance with contralateral angular gyrus in groups 1 and 2 with the addition of lateral temporal regions in group 3. More extensive bilateral lateral temporal and posterior cingulate covariance emerged in group 4. Somewhat unexpectedly, the DMN showed a peak voxel count in group 3 (Fig. 3B), resembling the pattern seen in the primary sensory and motor networks (Fig. 1B).

Statistical Validation of Observed SCN Trajectory Profiles.

Noting distinct domain-related SCN trajectories (Figs. 1–3), we sought additional statistical support for these observations by modeling peak and prune and cumulative growth trajectory patterns for each of our SCNs (Materials and Methods). These follow-up analyses identified brain regions whose seed-based covariance conformed to each of the two trajectory patterns across age groups (Table S1). Overall, primary sensory and motor seeds showed covariance with several regions that significantly followed a peak and prune covariance trajectory model but few regions that significantly followed a cumulative growth trend. Conversely, seeds for the language-related, social–emotional, and executive control networks significantly covaried with numerous regions fitting the cumulative growth model, but only one region in one network conformed to a peak and prune trajectory. Only the DMN defied these categorizations, possibly reflecting the complex architecture of this network (Discussion). Importantly, the observed network trajectory signatures (Figs. 1B, 2B, and 3B) proved robust to experimental variations in statistical thresholding (Fig. S1).

Laterality and Sex Effects on SCNs.

Recent studies suggest that structural hemispheric asymmetries may emerge during specific phases of child development (18, 19, 22, 24, 39). Although the present study was not designed to address laterality effects, we performed a preliminary investigation of these issues by generating SCNs and voxel counts using seeds contralateral to those first analyzed (Figs. 1–3). The contralateral seeds produced SCNs that largely mirrored our main findings, resulting in consistent SCN trajectory signatures (Fig. S2).

Reported sex differences in childhood brain maturation (40, 41) prompted us to consider similar differences in SCN development. Therefore, we divided each age group by sex, generated sex-specific SCN maps, and plotted corresponding voxel count trajectories for three SCNs, one from each of the three broad functional domains (Fig. S3). We chose the speech (L IFGo) and salience (R FI) networks in light of the view held by some authors that these cognitive domains may mature earlier in females (42–44), and we included the primary auditory [R Heschl's gyrus (HG)] network without specific a priori predictions. In general, voxel plot trajectory graphs (Fig. S3) depict earlier network expansion in females, particularly for speech (IFGo) and salience (FI) networks. Direct statistical contrasts between males and females at each age suggested more nuanced sex differences (Table S2), underscoring the need for future studies designed to address this issue. Overall, however, these data suggest subtle but compelling network topology and trajectory differences between males and females, consistent with the existing literature (40–42).

Subjects and MRI Methods.

Deidentified data were obtained from the National Institutes of Health (NIH) MRI Study of Normal Brain Development's Pediatric MRI Data Repository (33). This multisite study includes T1 MRI scans of 432 normal children between the ages of 4.5 and 21 y (no older than 18 y at enrollment). Further details regarding MRI methods, demographics, inclusion and exclusion criteria, and other technical specifications have been published (33). Briefly, MR images were acquired using standard 1.5 T MRI scanners (GE; Siemens). Whole-brain T1-weighted image datasets were acquired using standard techniques (time to repetition median = 23 ms, time to echo median = 10 ms, matrix median = 256 × 256 × 124, and median voxel volume = 1.32 mm³). We used all nonduplicate high-quality scans available in the database from right-handed subjects to select our study sample. Based on estimates of statistical power from previous structural covariance MRI (scMRI) studies (8), we constructed four sex-matched groups (n = 75, 34 males and 41 females) by first selecting the 75 oldest (group 4) subjects. We then selected the 75 youngest (Group 1) subjects, matching for sex. We divided the remaining subjects into two equally sized groups before selecting 75 nonduplicate subjects surrounding the respective group means, again matching for sex (groups 3 and 4). Resulting group characteristics are shown in Table 1.

Data Analysis.

We created customized image analysis templates by normalizing, segmenting, and averaging T1 images using SPM5 according to a recently proposed processing pipeline (57, 58). First, images were transformed into standard space using a 12-parameter affine-only linear transformation and segmented into three tissue classes representing gray matter, white matter, and cerebrospinal fluid. Smoothly varying intensity changes as well as artifactual intensity alterations as a result of the normalization step were corrected for using a standard modulation algorithm within SPM5. The resulting segmented maps were then smoothed using a 12-mm full-width at half-maximum Gaussian kernel. To study network structural covariance, we derived gray matter intensities using 4-mm-radius spherical seed regions of interest (ROIs) chosen using previously published ICN peak foci from the adult ICN literature and adjusted to our childhood template using relevant anatomical landmarks to avoid seed placement too near the pial surface. Seed ROIs were selected within right calcarine sulcus (primary visual cortex) (59), right Heschl's gyrus (primary auditory cortex) (60), right precentral gyrus (primary motor cortex) (61), left inferior frontal gyrus, pars opercularis (Broca's area) (14), left temporal pole (14), right frontoinsular cortex (14), right dorsolateral prefrontal cortex (8), and right angular gyrus (14). These regions anchor the visual, auditory, sensorimotor, speech, semantic, salience, executive control, and default-mode networks, respectively (4, 7, 8, 14). We performed separate condition-by-covariate analyses for each seed region in which the mean seed gray matter intensity was the covariate of interest and age group was the grouping variable. This design enabled us to determine the whole-brain patterns of seed-based structural covariance in each age group. Resulting correlation maps were thresholded at P < 0.001, corrected for FWE, and displayed on the normalized pediatric brain template to allow qualitative comparisons between age groups. To quantify differences in developmental trajectory across networks, we calculated the total number of significant ipsilateral, contralateral, and whole-brain voxels and plotted these by age group (Figs. 1–3).

Follow-up analyses sought statistical support for the domain-related voxel plot trajectories observed in Figs. 1B, 2B, and 3B by modeling the proposed primary sensory and motor trajectories as peak and prune (−1 − 1 + 3 − 1) contrasts and the language-related and cognitive trajectories as cumulative growth (−1 − 0.5 + 0.5 + 1) contrasts (Table S1). Both contrasts were applied to each seed to identify regions that significantly conformed to each model (P < 0.001, uncorrected). These analyses were inclusively masked by the largest group map (group 3 for peak and prune and group 4 for cumulative growth contrasts), with the mask thresholded more generously (P < 0.01, FWE-corrected) to avoid overly constraining the search volume.

Additional analyses using contralateral seeds (Fig. S2) were performed by changing the sign on each seed's x coordinate and following otherwise identical procedures to those used in generating the results shown in Figs. 1–3. Likewise, sex effects were examined by dividing each group by sex within otherwise identical statistical models. Sex-specific group SCNs were derived for voxel plot trajectories (Fig. S3) using a less-stringent statistical threshold given the lower sample sizes (P < 0.01, FWE-corrected). Linear group contrasts sought regional head-to-head sex differences in seed covariance at each age (P < 0.001, uncorrected, inclusively masked to the group average map at P < 0.05, uncorrected) (Table S2).