Developmental decorrelation of local cortical activity through adolescence supports high-dimensional encoding and working memory

Abstract

Adolescence is a key period for the maturation of cognitive control during which cortical circuitry is refined through processes such as synaptic pruning, but how these refinements modulate local functional dynamics to support cognition remains only partially characterized. Here, we used data from a longitudinal, adolescent cohort (N = 134 individuals ages 10–31 years, N = 202 total sessions) that completed MRI scans at ultra-high field (7 Tesla). We used resting state fMRI data to compute surface-based regional homogeneity (ReHo)—a measure of time-dependent correlations in fMRI activity between a vertex and its immediate neighbors—as an index of local functional connectivity across the cortex. We found widespread decreases in ReHo, suggesting increasing heterogeneity and specialization of functional circuits through adolescence. Decreases in ReHo included a spatial component which overlapped with sensorimotor and cingulo-opercular networks, in which ReHo decreases were associated with developmental stabilization of working memory performance. We show that decreases in ReHo are associated with higher intrinsic coding dimensionality, demonstrating how functional specialization of these circuits may confer computational benefits by facilitating increased capacity for encoding information. These results suggest a remodeling of cortical activity in adolescence through which local functional circuits become increasingly specialized, higher-dimensional, and more capable of supporting adult-like cognitive functioning.

1. Introduction

Adolescence is a significant period of maturation for brain systems supporting executive functioning, during which cognitive performance improves and stabilizes, becoming more accurate and less variable over time (Tervo-Clemmens et al., 2023). This refinement of cognitive performance is thought to arise from a cascade of neurobiological changes that occur primarily within the brain’s higher-order association cortices, including the prefrontal cortex (PFC). These developmental changes include an acceleration in synaptic pruning in PFC into adulthood (Petanjek et al., 2011), refinements in the excitation-inhibition balance of neurotransmitter systems (McKeon et al., 2024b, Perica et al., 2022), and increases in intracortical myelination (Paquola et al., 2019a, Whitaker et al., 2016). Refinement of the synaptic organization and local connectivity of developing circuits has the potential to dramatically reshape functional interactions in the cortex. However, while age-dependent changes in large-scale functional connectivity have been extensively characterized, changes in the properties of local functional coupling through the adolescent period are less well understood.

Despite consistent findings that functional connectivity (FC) shifts through adolescence (Keller et al., 2023, Marek et al., 2015), characterizations of developmental change in whole-brain and network-level FC through adolescence have produced divergent results. Recent work has attempted to identify global organizational principles governing FC development, including grounding the maturation of functional networks in a sensorimotor-to-association (S-A) axis, whereby somatomotor regions tend to show strengthening FC while association cortices show decreases in FC (Luo et al., 2024, Sun et al., 2024). Emergent work leveraging large, multi-study samples has demonstrated an overall shift toward increased functional segregation through adolescence, though with substantial variation in the magnitude and tempo of this effect across brain regions (Sun et al., 2024). Existing literature has thus identified adolescence as a time of substantial reorganization of the functional connectome, during which brain networks specialize and refine to support the changing cognitive demands on the cortex during the transition to adulthood.

While age-related changes in long-range network connectivity have been frequently addressed, less work has characterized changes in local (intra-regional) FC, which captures the local synchronization of spontaneous neural activity. Whereas differences in long-range FC have been linked to white matter properties such as axonal myelination (Benamer et al., 2020, Seidl, 2014, Sturman and Moghaddam, 2011), connectivity at local scales (e.g., 1–2 mm) may more closely reflect mechanisms which affect local circuit properties within gray matter. This may potentially include the extent of synaptic interconnectedness of neighboring circuits (and pruning of these synapses), intracortical myelination of excitatory and inhibitory axons within gray matter (Micheva et al., 2016), and homogeneity of cell types (Jiang and Zuo, 2016). Each of these microstructural mechanisms contributes to the density and/or strength of connections within local gray matter circuitry and thus local cortical synchrony, and their maturation may therefore be reflected as developmental changes of local functional circuit characteristics

A key approach for characterizing local FC from BOLD fMRI data is the computation of regional homogeneity (ReHo). ReHo provides a measure of local FC at the millimeter scale, reflecting the integration of neuronal activity over small cortical distances via the functional coherence of a given area (i.e., voxel or vertex) with its nearest neighbors (Zang et al., 2004). Prior work has demonstrated that regional variation in ReHo correlates to the relative functional complexity of different brain regions arising from differences in cytoarchitectonic heterogeneity and regional synaptic wiring, particularly for differences between unimodal, multimodal, and transmodal areas (Song et al., 2014). Accordingly, ReHo may reflect the functional properties of a region within the hierarchical organization of the brain. Mapping of interindividual ReHo variability has revealed covariance networks that are distributed systematically across common brain networks (Jiang and Zuo, 2016).

Individual differences in regional homogeneity have also been linked to age-related functional development. Initial studies of ReHo spanning late childhood through early adulthood have suggested developmental changes present across a wide array of brain regions (Lopez-Larson et al., 2011) and networks (Hong, 2023), with changes in ReHo nearly ubiquitously decreasing, though with substantial variability in the magnitude of decreases across brain regions. These adolescent changes in ReHo may represent a process which continues through the lifespan (Wu et al., 2007). ReHo alterations have also been implicated in a number of mental health disorders which emerge or strengthen during adolescence. Decreases in cortical ReHo have been observed in schizophrenia beginning in adolescence (Shan et al., 2021, Yang et al., 2022), as well as major depressive disorder (Li et al., 2014), and bipolar disorder (Xu et al., 2019), with suggestions that interventions may lead to longitudinal increases in ReHo across conditions (Lu et al., 2023, Shan et al., 2021). However, the brain regions implicated in these studies have thus far been highly heterogenous, and consideration of normative changes in ReHo are critical for understanding the spatial and temporal patterning of these effects. Further, although preliminary evidence has associated developmental changes in ReHo with general cognitive function (Hong, 2023), associations between ReHo and cognitive abilities in adolescence remain largely unknown. Finally, investigations into how ReHo might affect computational capacities of a circuit to support enhanced cognitive processing remain sparse. Here, we hypothesized that decreases in local FC will support improved cognitive performance by increasing the computational capacity of cortical functional circuitry. To address this, we leverage a large, adolescent cohort with ultra-high field (7 Tesla) MRI and fMRI data to demonstrate dissociable patterns of ReHo which support improvements in the stability of working memory performance. Additionally, we show that developmental decreases in ReHo are associated with an increase in the functional dimensionality of BOLD signals within cortical regions, suggesting that functional specialization of neural circuitry facilitates an increase in coding efficiency, which may support the emergence of adult levels of cognitive function.

2. Methods and materials

2.1. Participants

Data from 134 typically developing participants were analyzed (52.5 % assigned female at birth, ages 10–31 years). Participants completed between 1 and 3 visits as part of an accelerated longitudinal design, with longitudinal visits scheduled approximately 18 months apart. In total, 203 individual scans were analyzed (n = 134 initial visits, with n = 52 visit 2 follow-ups, and n = 17 visit 3 follow-ups). Participants were recruited from the local Pittsburgh population and were screened for psychiatric illness and MRI contraindications. The study received approval from the University of Pittsburgh’s Institutional Review Board (IRB). Participants and their parents (for those under 18) provided informed consent, and participants were compensated for their participation.

2.2. Working memory task

Working memory was assessed using a memory guided saccade (MGS) task, which has been reported on in prior studies from our group (McKeon et al., 2024b). This task involved an oculomotor paradigm in which participants first fixated a central dot, then shifted fixation to a peripheral cue. Once the cue disappeared, participants returned their gaze to the central fixation point and fixated for a variable delay epoch (6–10 sec). When the fixation target was extinguished, participants performed a memory guided saccade to the recalled location of the previous cue. Data were acquired during a separate session from the MRI, during which EEG data was also collected from participants. Eye movements were assessed based on electrooculogram (EOG) channels placed on the face to measure muscle activity associated with eye movements. Eye positions were calibrated prior to the MGS task by asking participants to fixate on a series of dots placed along the horizontal meridian. Performance was assessed by measuring the accuracy of the memory guided saccade (degrees of visual angle error between the memory guided saccade and the true target location), as well as the latency to initiate both the visually guided and memory guided saccades. Trial-to-trial variability in each measure was computed as the standard deviation of each measure across trials.

2.3. MR data acquisition

MR data were acquired at the University of Pittsburgh Medical Center Magnetic Resonance Research Center on a Siemens 7 Tesla scanner. Structural images were collected using an MP2RAGE sequence (1 mm isotropic resolution, TR = 6000 ms, TE = 2.47 ms, flip angle 1 = 4°, flip angle 2 = 5°, voxel size = 1 ×1×1 mm; total duration = 5 m 14 s). Resting-state fMRI data was acquired using a 3D T2 * -weighted scan with TR = 2.04 s, TE = 23 ms, voxel size = 2 × 2 × 2 mm, with a scan duration of 8 m 3 s. During the rs-fMRI scan, participants were presented a blank screen and were instructed to keep their eyes open, remain still, and to stay awake.

2.4. MR and fMRI preprocessing

Unified structural images (UNI) from the MP2RAGE acquisition were corrected for residual B1 bias field inhomogeneities using the unified segmentation algorithm from SPM (Ashburner and Friston, 2005, Weiskopf et al., 2011). This was implemented using a containerized version of SPM12 (https://hub.docker.com/r/bids/spm/tags) with default parameters. Following bias correction, FreeSurfer's longitudinal processing stream was run on UNI images (including cross-sectional, template, and longitudinal steps) using the containerized FreeSurfer BIDS app version 7.4.1 (https://hub.docker.com/r/bids/freesurfer/tags). Mean values for each atlas region of interest (ROI) were computed per subject by transforming data from the fsaverage template into session-specific cortical surfaces using FreeSurfer’s spherical registration-based surface resampling algorithm.

Corrected images were pre-processed with fMRIPrep 23.2.1 (Esteban et al., 2019, Esteban et al., 2018), which is based on Nipype 1.8.6 (Gorgolewski et al., 2011, Gorgolewski et al., 2018), to create surface-based templates and registrations for aligning functional data. Functional images were post-processed using the eXtensible Connectivity Pipeline-DCAN (XCP-D) (Mehta et al., 2023). XCP-D provides a robust pipeline designed to provide rigorous motion correction and data processing. Briefly, we used the ‘36 P + despike’ pipeline described by Ciric et al. (Ciric et al., 2017), which included a 36-parameter nuisance regression including 6 motion dimensions, their quadratic expansion, as well as global, white matter, and CSF signals and their derivatives. Temporal spikes in the BOLD data were removed using AFNI’s 3dDespike, and time courses were bandpass filtered between 0.01 and 0.1 Hz. This approach allows for removal of motion-related signal spikes without reducing the number of timepoints available, which could artificially constrain dimensionality estimates (based on the methods described below). A more detailed automated methodological description for fMRIprep and XCP-D processing steps is available in the Supplemental Material. We also show a comparison of the 36 P + despike pipeline to 36 P + scrub, in which high motion time points are removed entirely, on ReHo estimates. Briefly, these pipelines produced highly similar estimates of ReHo (r = 0.89), with highly conserved developmental effects. Scrubbing led to slightly higher estimates of ReHo than despiking for cases with high amounts of motion, such that despiking produced slightly more conservative estimates of age effects overall.

2.5. Surface-based ReHo

Surface-based ReHo values were computed as part of the XCP-D processing pipeline as described above and in the supplemental material. Briefly, pre-processed fMRI BOLD time courses were mapped to the cortical surface mesh, and 2D ReHo was computed using the surface-based 2dReHo function (Zhang et al., 2019, Zuo et al., 2013). For each vertex within the gray matter ribbon, the degree of correspondence with the 6 nearest neighbor vertices was computed as the Kendall’s coefficient of concordance (KCC). This allowed ReHo to be computed entirely within gray matter voxels while following the contour of the cortical surface. Previous empirical studies have shown that this approach produces more robust estimates for connectomic analyses such as ReHo (Zuo and Xing, 2014), since typical 3D approaches are confounded by partial voluming effects (i.e., including the correspondence between gray and partially-white matter voxels in ReHo estimates), or across cortical folding patterns (e.g., from one gyrus to the next, which may be close in Euclidian distance, but far apart when following the cortical surface). To address this, in the Supplemental Material, we additionally performed direct comparisons of 2D surface-based estimations of ReHo with 3D volumetric estimates and show that values are correlated across the whole brain, and produce comparable age effects. Finally, vertex-wise ReHo maps were parsed using the connectome workbench (Marcus et al., 2011) to extract regional mean ReHo values for each cortical region from the Human Connectome Project multimodal parcellation (HCP-MMP) (Glasser et al., 2016). Temporal SNR (tSNR) was estimated from the fMRI data as the ratio of BOLD signal mean to standard deviation to assess signal quality; regions of low tSNR (gray regions in Fig. 2A), primarily in the orbitofrontal cortex and temporal pole, were excluded from further analyses.

Fig. 2.

Spatial topography and age-related change of ReHo. (A) Mean ReHo map by cortical region, averaged across all subjects. (B) Cognitive decoding of the mean ReHo map through correlation with 124 cognitive term maps derived from NeuroSynth. Positive correlation values indicate strong alignment with high-ReHo values, while negative correlations indicate alignment with low-ReHo regions. (C) Age-related change in ReHo by cortical region. Filled regions indicate FDR-corrected significant age-related change, with blue values indicating decreases in ReHo with age. (D) Histogram of regional F-values of the age-related change term from a GAM model (with sign assigned based on a linear mixed effect model to capture the direction of change with age). Blue bars indicate FDR-corrected significant change. (E) Change in whole-brain ReHo with age, controlling for head motion and sex. (F) Network assignments of regions with significant age-related change (proportion of regions with FDR-corrected significant change in each network) based on the Yeo 7-network parcellation (Thomas Yeo et al., 2011).

2.6. Intrinsic dimensionality estimation

In order to demonstrate potential computational correlates of changes in local functional connectivity, we leverage approaches to characterize the complexity of neuronal activation within brain regions based on their intrinsic functional dimensionality. We used a measure termed representational dimensionality (RD), which has previously been applied to BOLD fMRI data (Sheng et al., 2022). RD is a PCA-based approach that estimates the number of meaningful, non-noise signal components present within a given brain region as a measure of the region’s computational capacity. Briefly, for each region within the cortical surface, time courses of each vertex containing gray matter BOLD signal were used as the basis for a PCA using the R function ‘prcomp’. Eigenvalues for each principal component were computed, and RD (also previously referred to as effective RD, or RD_eff) was calculated as the number of eigenvalues greater than 1, divided by the proportion of the total variance explained by those eigenvalues. Individual differences in RD have been proposed to reflect differences in the ability to efficiently encode information by representing multiple spatial patterns within a given brain region, and therefore differences in encoding dimensionality (Sheng et al., 2022).

2.7. Statistical analyses

We used generalized additive models to assess age-related change and for comparisons between measures (e.g., ReHo-vs-performance, ReHo-vs-dimensionality), both at the whole-brain and per-region of interest (ROI) level. The primary motivation for this was to allow for non-linear effects of age, as well as non-linear effects of potentially confounding covariates. Head motion is known to have a strong contribution to ReHo, and while our preprocessing pipeline was designed to mitigate this, we control for mean framewise displacement (FD) in all analyses to limit residual effects of motion on the reported effects. Additionally, surface-based preprocessing approaches generally produce uneven vertex spacing, which can result in differences in the true spacing between neighboring vertices across participants and brain regions, directly biasing time course analyses such as short-distance functional connectivity which have strong distance dependence (Jeganathan et al., 2024). Given that surface area has well-known developmental variation (Bethlehem et al., 2022), it is likely that there are differences in the degree of surface compression onto individual vertices between younger and older participants. Thus, we include Freesurfer-based estimates of each ROI’s surface area normalized by the number of vertices within each region (i.e., surface area per vertex, or the surface area compression, see Supplemental Material) as covariates in all analyses to control for individual and regional differences. The GAM framework allows us to incorporate these covariates without assumptions of the linearity of any effects. GAM models further incorporated random effects of participant ID to account for the repeated longitudinal observations. Inspection of results, including preliminary interaction analyses when possible, did not indicate strongly lateralized patterns of results, so regional analyses were performed by combining corresponding ROIs across hemispheres and controlling for a main effect of hemisphere in statistical models. For visualization purposes, regional effects are presented on lateral and medial views of a left hemisphere brain image, using the R library ggseg (Mowinckel and Vidal-Piñeiro, 2020). Statistical outliers for all ROI-based analyses were identified as data points more than 2 standard deviations from the mean and were removed prior to analysis. GAMs were fit with REML estimation using the mgcv library in R (Wood, 2006). ReHo was always included as the independent variable, and smooth terms were used for age, head motion and surface area compression, limiting the k value to < =3. This limit acted as an additional regularization parameter on the GAM fits which helped prevent overfitting, e.g., by constraining the number of age-related inflection points, since prior findings in adolescent cognitive developmental studies have suggested relatively low dimensional non-linearities (Tervo-Clemmens et al., 2023). Inspection of fits indicated that a greater number of knots overall did not tend to improve model fits, and we note that this did not preclude estimation of fewer knots (i.e., more linear fits). When comparing brain (dimensionality) or behavioral (MGS task performance data) measures to ReHo, linear terms were used within the GAM framework since we did not have a priori expectations of non-linear effects for these associations.

3. Results

3.1. Working memory accuracy and stability increase during adolescence

Consistent with our previous findings from an overlapping cohort (McKeon et al., 2024b), we found significant age-related improvements in both accuracy and latency measures from the memory guided saccade (MGS) task. Improvements in mean accuracy (degrees of visual angle error, F=14.9, p = 1.1 ×10⁻⁶, p_bonf=6.7 ×10⁻⁶, Fig. 1A) as well as trial-to-trial variability in accuracy (F=10.6, p = 4.9 ×10⁻⁵, p_bonf=2.9 ×10⁻⁴, Fig. 1D) were evident. As expected, mean latency to initiate a memory guided saccade (F=6.0, p = 0.0033, p_bonf=0.020, Fig. 1C) and trial-to-trial variability in time to initiate both the visually guided saccade (F=5.7, p = 0.004, p_bonf=0.025, Fig. 1E) and the memory guided saccade (F=22.7, p = 2.7x10⁻⁹, p_bonf=1.6 ×10⁻⁸, Fig. 1F) decreased significantly with age. We note that while the visually guided saccade was performed as part of the working memory task, it refers to the initial saccade made as subjects viewed and encoded the target, and thus represents a reflexive oculomotor response more likely related to processing speed and attention and not working memory maintenance or retrieval per se. Since variability in behavioral measures frequently scales with the mean performance level, we performed an additional sensitivity analysis to determine if the coefficient of variation (CV; standard deviation divided by mean) for each of the trial-to-trial variability measures also changed with age. This identified significant Bonferroni-corrected age-related decreases in the trial-to-trial CV of the latency to initiate a memory guided saccade (F=12.7, p < 9.8x10⁻⁶, p_bonf=8.8x10⁻⁵), as well as decreases in VGS latency CV that did not survive correction (F=4.4, p = 0.013, p_bonf=0.13), but no effect of age on trial-to-trial CV of accuracy (F=1.64, p = 0.2), consistent with previous reports for this task in a different cohort (Montez et al., 2017). These behavioral findings add strong evidence that working memory improves, and is engaged more stably, from early adolescence to early adulthood.

Fig. 1.

Performance on the memory guided saccade task as a function of age. Top row: Mean performance metrics per participant across all trials, including (A) accuracy of the memory guided saccade (degrees of visual angle error from the actual target location), (B) latency to initiate the visually guided saccade, and (C) latency to initiate the memory guided saccade. Bottom row (D-F): Trial-to-trial variability in each of the measures, computed as the standard deviation across trials per subject.

3.2. Developmental refinements in regional homogeneity

Mean regional homogeneity (ReHo) was computed across participants for each region to provide an overall assessment of its spatial distribution across the cortex (Fig. 2A). This revealed a pattern with highest ReHo in frontoparietal regions, as well as regions near the posterior cingulate cortex, indicating regions with a high degree of local functional similarity. Conversely, low ReHo was observed in the temporal cortex, as well as in sensorimotor regions. To characterize these patterns in terms of their relevance to psychological functions, we computed the spatial correspondence between the across-cortex ReHo map and cortical maps derived from Neurosynth (Yarkoni et al., 2011) based on 123 common behavioral terms. Specifically, based on prior approaches (Hansen et al., 2024, Shafiei et al., 2020, Sydnor et al., 2024), meta-analytic maps for terms spanning a broad range of psychological processes were derived from Neurosynth data using the Neuroimaging Meta-Analysis Research Environment (NiMARE) (Salo et al., 2023). Meta-analytic term maps were mapped to fsLR space and correlated to ReHo maps, which were corrected for differences in surface area compression (see Supplemental material), to identify associations between spatial patterns of ReHo and topography of activation associated with each term (Fig. 2B). Inspection of the highest correlations from this analysis identified strong positive associations between the cortical ReHo map and terms associated with working memory (‘manipulation’, ‘working memory’), attention (‘attention’), and decision making (‘uncertainty’, ‘task difficulty’), indicating that these psychological functions recruit brain regions with relatively higher ReHo. Conversely, negative correlations were identified with terms associated with emotional processing (‘fear’, ‘emotion’), and language (‘communication’, ‘morphology’), indicating high alignment of these terms’ cognitive maps with low-ReHo regions.

Whole-brain ReHo, averaged across all cortical regions for each participant session, showed significant age-related changes (F=24.7, p = 2.17x10⁻⁶) while controlling for non-linear effects of head motion (mean FD), cortical surface area per vertex, sex, and random effects of participant (offset). Inspection of the marginal effect of age on ReHo (Fig. 2E) indicated marked and persistent decreases in ReHo with age that continued linearly throughout the age range of the cohort. To understand the spatial extent of developmental declines in ReHo, we repeated this analysis for each cortical region. We identified relatively widespread, significant age-related decreases in ReHo after FDR correction (Fig. 2C). For visualization of the magnitude and direction of significant regional effects, F-values are shown for cortical regions with a sign assigned (based on the sign of the age term from a separate linear model) to capture whether age-related change is generally increasing or decreasing. Inspection of a histogram of age effects revealed nearly ubiquitous decreases in ReHo, with over half of the assessed regions (83/159, 52.2 %) surviving FDR-correction (Fig. 2D). Age effects were distributed across functional networks, including substantial representation in the somatosensory (22.9 %), default mode (21.1 % of regions with age effects), dorsal attention (18.8 %), frontoparietal (16.9 %) and ventral attention (12.7 %) networks, and visual (8.4 %) networks, while excluding the limbic network (0 %) (based on the Yeo 7-network parcellation (Thomas Yeo et al., 2011) as applied to Glasser parcellated data (Paquola et al., 2019b)).

3.3. Spatial components of cortical ReHo variation

Given the widespread age effects identified for ReHo across the cortex, along with positive correlations of ReHo values across regions (see Supplemental Figure 4) and a significant decrease in ReHo with age at the whole-brain level, we next sought to identify dominant spatial components of ReHo variability and their refinement with age. Specifically, we sought to identify global processes by which ReHo decreases across spatially distributed systems. To address this, we computed a PCA across regions and participants to identify distinct spatial components contributing to the observed cortical ReHo pattern. Based on cross-validation based methods for approximating the number of meaningful components (Josse and Husson, 2012), we identified 5 principal components (PCs), which captured the majority of variability in ReHo across the cortex. For each of these 5 components, we computed factor scores per participant (the degree to which each subject expressed the component in their own ReHo data) and computed the correlation of each component score with age. This procedure identified two PCs which changed significantly with age after controlling for non-linear effects of head motion, as well as sex, and random effects of participant: PC1 (15.05 % of total ReHo variance explained; F_age=22.4, p = 6.0x10⁻⁶, p_bonf=3.0x10⁻⁵) and PC2 (5.35 % of total ReHo variance explained; F_age=11.3, p = 9.6x10⁻⁴, p_bonf=0.0048). PCs 3–5 did not show any significant associations with age (p_uncorrected>0.3).

Inspection of the factor loadings indicated that PC1 contained strong contributions from somatomotor regions as well as a significant number of fronto-parietal regions (Fig. 3A). Network contributions were well distributed among the dorsal attention (DA) (25.0 %), somatomotor (19.4 %), default mode (18.1 %), visual (13.9 %), frontoparietal (12.5 %), and ventral attention (11.1 %) networks, while excluding limbic regions (0 %). ReHo values from the ROIs comprising the top 20 % of loadings in each component were highly correlated to the factor scores for each component (see Supplemental Figure 5), reflecting that factor scores can be interpreted as representing a weighted mean of ReHo values from across these prominent ROIs. Analysis of the factor scores indicated the age-related decreases in PC1 ReHo were protracted, persisting through the 20’s (Fig. 3B). PC2 similarly showed strong involvement of sensorimotor regions near the motor cortex, along with regions near the insula, and a significant contribution of dorsomedial prefrontal regions (Fig. 3D). Network contributions were dominated by the somatomotor (50.0 %) and ventral attention (38.9 %) networks, with strong qualitative overlap with the cingulo-opercular network as defined in (Gordon et al., 2016, Power et al., 2011). Inspection of age-related change in ReHo of this component appeared to occur more rapidly early in adolescence, before showing a more prominent plateau through the 20’s compared to PC1 (Fig. 3E).

Fig. 3.

Spatial components of ReHo for the first (top row) and second (bottom row) principal components. (A, C) Loadings of the first (PC1) and second (PC2) principal components of ReHo across the cortex. Regions with the top 20 % of loadings are illustrated. (B, D) Age-related change in PC1 and PC2 factor scores, showing protracted developmental decreases. (E) Association of PC2 with variability in the time to initiate a memory guided saccade across trials. Lower ReHo values were associated with reduced trial-to-trial variability (PC2).

3.4. ReHo and maturation of working memory performance

In order to assess the association of ReHo with cognitive performance, we compared factor scores for PC1 and PC2 to the six performance measures from the MGS task (Fig. 1). This identified two associations that survived Bonferroni correction after controlling for head motion and surface area compression: both PC1 and PC2 (t = 3.14: p = 0.002: p_bonf=0.012, Fig. 3E) were significantly associated with trial-to-trial variability in MGS latency. This effect persisted when controlling for age for PC2 (t = 2.07, p = 0.04), but not PC1 (p > 0.1). The direction of the observed association was consistent with the identified age effects: reductions in ReHo were associated with improved task performance (reduced trial-to-trial variability).

3.5. Intrinsic dimensionality increases through adolescence

Developmental decreases in ReHo reflects increased functional segregation of local circuits that arises from a decorrelation of signals. However, while our results suggest a link between decreases in ReHo and maturation of cognitive performance, an open question is what computational change is either induced from, or reflected in, the developmental decorrelation of circuit activity. A likely possibility is that decreases in the correlation among proximal voxels results in increases in the spatial dimensionality of a region by producing a greater number of distinct neuronal signatures, reflecting an increase in the information content coded by cortical regions. To demonstrate the possibility that this relationship between local correlation, as indexed by ReHo, and the dimensionality of a region reflects the computational benefit conferred by functional specialization, we computed an intrinsic dimensionality estimate (IDE) for each cortical region based on an index of effective representational dimensionality (RD) introduced in prior work (Sheng et al., 2022). RD is a PCA-based approach that attempts to estimate the number of meaningful (non-noise) signals represented spatially within each region. The mean RD across the cortex is shown in Fig. 4A.

Fig. 4.

Estimates of intrinsic dimensionality based on representational dimensionality (RD) estimates. (A) Mean RD by region across the cortex. (B) Statistical association of RD with age after controlling for head motion. Red indicates statistically significant increases in RD with age. (C) Whole brain RD (averaged across all cortical regions) increases with age. (D) Widespread negative associations between RD and ReHo after controlling for age and head motion. (E) Association between ReHo and RD based on whole-brain estimates.

To characterize the correspondence of RD to developmental changes in ReHo, we first computed age-related change in dimensionality for each region (Fig. 4B) and at the whole brain level (Fig. 4C). We found that whole-brain RD increased through adolescence after controlling for head motion and surface area compression (F=6.09, p = 0.014). Inspection of FDR-corrected regional age effects revealed a pattern of nearly exclusively increases in dimensionality, spanning parietal, ventral occipital, and both lateral and medial prefrontal regions. RD showed strong, consistent associations with ReHo, both regionally (Fig. 4D) and at the whole brain level (F=8.07, p = 0.0050, Fig. 4E), with negative associations reflecting that relatively lower levels of ReHo were associated with higher dimensionality even after controlling for age. This pattern of results was highly robust across two other measures of dimensionality tested (see Supplemental Material and Supplemental Figure 6). Collectively, these findings suggest that age-related decreases in locally correlated activity are coupled to robust increases in the dimensionality of functional signals.

4. Discussion

Characterizing changes in the functional architecture of the cortex through adolescence has been a key aspect of understanding how developmental changes in cortical properties contribute to the maturation of executive functioning. Using resting state fMRI data from a large cohort of adolescent and young adult participants scanned at ultra-high field (7 T) MRI, we found widespread increases in the functional segregation of cortical circuit activity at the millimeter scale, indexed by fMRI signals becoming more independent across immediately adjacent cortical areas, with age. For the purposes of these results, we operationalize local circuitry as referring to connectivity among neuronal populations within gray matter at the millimeter scale, which can be detected through interactions among neighboring voxels. This result is consistent with prior work in adolescence (Hong, 2023, Lopez-Larson et al., 2011) and across the lifespan (Wu et al., 2007), suggesting that with age, functional processing tends towards functional specialization. Our results suggest that this increased specialization allows brain regions to code an increasing number of independent signals, potentially increasing coding capacity and efficiency, facilitating developmental improvements in cognitive control.

Recent studies of whole-brain functional connectivity have suggested a developmental trend towards increased segregation of long-range functional connections, with protracted increases in segregation through adolescence apparent across ventral attention, frontoparietal and default mode networks (Sun et al., 2024). The results presented here suggest a similar process occurring at local scales: wide-spread decreases in functional coupling among neighboring vertices were observed across most cortical regions, including those in the somatomotor, dorsal and ventral attention, frontoparietal, and default mode networks. Comparison to volumetric ReHo (Supplemental Material) revealed similar age-related decreases present across a range of neighborhood sizes. Together with increased segregation among long-range functional connectivity (Sun et al., 2024), these results indicate that increasing functional specialization may be occurring across multiple spatial scales through adolescence, suggesting consistent shifts toward functional specialization. Importantly, while head motion can induce strong biases in ReHo estimation due to the spatial blurring of BOLD signal across nearby voxels, developmental effects persisted after both rigorous correction for head motion during preprocessing, and statistically controlling for non-linear effects of head motion in regression analyses. Overall, these results suggest an increase in the functional specialization of cortical circuitry, as activity-related signals become more functionally independent.

In the context of brain networks, increased segregation has been observed for both functional (Sun et al., 2024) and structural (Baum et al., 2017) connections, supporting increased modularity of large-scale cortical circuitry. Although protracted changes in white matter microstructure (Lebel and Beaulieu, 2011, Simmonds et al., 2014) and increases in myelination (Morris et al., 2020) may contribute to developmental shifts in these long-range connections, they are less likely to mediate the increases in segregation and specialization we report for local connectivity which likely implicates circuitry localized within gray matter. Instead, these findings may point to underlying mechanisms rooted in regional intracortical wiring properties (Park et al., 2022). Synaptic connectivity increases dramatically in early childhood (Huttenlocher and Dabholkar., 1997), leading to an over-proliferation of connections (Rakic et al., 1994) which are ultimately pruned through adolescence (Petanjek et al., 2011). Synaptic pruning, through a Hebbian process (Faust et al., 2021), retains synapses that are necessary for adult-like processing, which may support improved cognitive function and generalization of task learning (Averbeck, 2022). Adolescent pruning of synapses leads to sparser but more reliable connectivity at local spatial scales, providing a potential mechanism for the decreases in ReHo we have observed. However, continued validation of imaging metrics including ReHo based on electrophysiological and histological data are critical for understanding the precise mechanisms contributing to changes in local FC at these spatial scales, and for appropriate interpretation of these results.

Comparison of developmental changes in ReHo with changes in the intrinsic dimensionality of the BOLD signal demonstrates a potential computational benefit of increased functional specialization which may facilitate the maturation of cognitive processes. Decoupling of nearby vertices allows more identifiable signal components to be represented within individual regions of the cortex, providing the opportunity to encode a greater diversity of signals within a given brain region. Importantly, while PCA-based methods will necessarily identify a fixed number of components (determined by the number of time points and/or vertices within a region), methods such as the representational dimensionality (RD) attempt to determine how many of these contain meaningful, non-noise signals. Low dimensional activity is highly robust to noise (Gallego et al., 2018, Tavares et al., 2015), which may make it an ideal coding strategy at younger ages when spontaneous neural activity is high and can interfere with task-related signals (Sydnor et al., 2023). However, low dimensionality limits the amount of information which can be encoded. We speculate that developmental reductions in spontaneous neural firing and corresponding increases in cortical signal-to-noise ratios (Fagiolini and Hensch, 2000, McKeon et al., 2024a, Toyoizumi et al., 2013) may allow signals to be encoded with greater fidelity. As signal fidelity increases, the need for redundant coding may decrease, allowing the same amount of information to be encoded in sparser neuronal populations while maintaining the robustness of these signals to noise. Sparse coding could allow neighboring neural populations to code distinct pieces of information, consistent with the decreases in ReHo we have observed, and resulting in computational benefits in terms of the dimensionality and complexity of neural signals across a cortical region. Such increases in the efficiency of neural signaling and complexity of neural signal processing may be critical for supporting complex cognitive processing and cognitive flexibility known to develop through adolescence (Fusi et al., 2016, Parr et al., 2024). Findings of increased segregation of long-range FC in conjunction with decreased ReHo may indicate that this shift towards sparse coding, allowing for increased functional specialization and information coding capacity, may represent a generalized mechanism of adolescent functional maturation which is present across spatial scales.

While decreases in ReHo are likely to reflect a greater diversity of neuronal signals within an ROI, this need not necessarily increase the true dimensionality of cortical regions. In the extreme case, voxels that are fully independent would result in a large number of distinct components, but these would likely not rise above the noise baseline, since each would be sparsely represented. Instead, our results indicate that specialization of neighboring voxels was associated with a greater array of non-noise sources within each ROI, potentially reflecting spatially constrained patterns of developmental change, consistent with recent neural recording results from developmental non-human primate studies which have found increases in temporal encoding dimensionality in PFC through adolescence (Zhu et al., 2024). One possible scenario for this result is an increase in the heterogeneity of functional modules, in which a large number of overlapping components are present within the ROI, but in which each has relatively weak associations among neighboring voxels. Such a result could be driven by mixed-selectivity (Rigotti et al., 2013), in which neurons (or groups of neurons, voxels, etc) participate in multiple computations, and computational networks, across time. This would result in a large number of spatial components, or dimensions, but with relatively weak local connectivity, since voxels would not be consistently interacting with the same neighbors. However, further work to develop and validate methods to assess spatial dimensionality, and extensions of fMRI studies to incorporate aspects of temporal dimensionality across both extended periods of time and different cognitive states, are important for understanding the extent to which existing methods can accurately distinguish noise and non-noise states.

While decreases in ReHo were observed across significant portions of the cortex, spatial decomposition of these effects suggested two distinguishable developmental components with distinct cognitive associations. The first component encompassed regions of the somatomotor, dorsal attention, and frontoparietal networks, but was not strongly correlated with any performance measures from our working memory task after controlling for age. On the other hand, a second component which implicated regions of sensorimotor and ventral attention networks showed rapid decreases in ReHo through adolescence, which decelerated through late adolescence and plateaued in the 20 s. This component showed an association specific to the trial-to-trial variability in initiating memory guided saccades. Notably, the developmental trend for this ReHo component mirrored that of the behavioral trajectory for MGS latency variability (Fig. 1F), which showed rapid decreases through adolescence before plateauing in young adulthood. Heightened intra-individual variability when executing cognitive tasks is a hallmark of adolescent (as compared to adult) executive function, with studies consistently finding development shifts toward greater behavioral stability as individuals mature towards adulthood (MacDonald et al., 2006, Tamnes et al., 2012, Williams et al., 2005). Such age-related reductions in variability have been linked to stabilization of functional brain states (Montez et al., 2017) and integrity of white matter tracts (Tamnes et al., 2012). Associations between ReHo and variability in initiating memory guided oculomotor responses in our working memory task overlapped substantially with the cingulo-opercular network (CON) (Gordon et al., 2016, Power et al., 2011) (recently termed the action mode network, or AMN (Dosenbach et al., 2024)), as well as sensorimotor regions near the motor cortex. Recent findings have suggested that the CON/AMN interacts with the newly identified somato-cognitive action network (SCAN) (Gordon et al., 2023) to implement and control goal-directed movement (Dosenbach et al., 2024), supporting the established role of the CON in exerting top-down control to regulate motor (Lu et al., 1994) and cognitive (Dosenbach et al., 2008, Dosenbach et al., 2006) responses. Our findings of specialization of the CON and somatomotor networks facilitating developmental decreases in the variability to initiate memory guided saccades thus support a model whereby high-fidelity of the CON and somatomotor network regions provide more reliable top-down regulation for initiating goal-directed actions. Although we were unable to interrogate the SCAN network directly, the presence of somatomotor and CON/AMN network regions within the same component may indicate a shared mechanism underlying maturation, perhaps mediated by the highly interconnected SCAN network, to facilitate these improvements. Future work leveraging prolonged acquisitions and precision functional mapping approaches (Demeter and Greene, 2024) would allow a more direct test of this hypothesis. Behaviorally, the lack of associations with accuracy are in accord with previous findings (McKeon et al., 2024b), and may reflect that regional neural computations support efficacy of generating an executive response that is present early in development, while accuracy may depend on long range integration of specialized regions (Nagy et al., 2004, Simmonds et al., 2014, Tamnes et al., 2013).

Together, these results provide evidence for the persistence of specialization across the cortex through adolescence, as well as identifying its possible role in increasing efficient information coding by supporting increased information complexity. Continued efforts to characterize the neural basis of local FC are critical to precisely understand how physiology of intracortical circuits is reflected in measures such as ReHo, including understanding the relative contribution of processes such as synaptic pruning, shifts in excitation:inhibition balance, and intracortical myelination. Establishing these links will help elucidate how deviations in the maturation of local FC between typically developing and clinical populations may reflect underlying differences in cortical physiology, particularly in the case of adolescent-emerging mental health disorders (e.g., schizophrenia, mood disorders, substance use disorders). The results presented here provide a foundation for understanding how changes in circuit dynamics contribute to the maturation of cognitive function—a prerequisite for elucidating how developmental disruptions in functional architecture may contribute to impaired neurocognitive development.

CRediT authorship contribution statement

Parr Ashley C: Writing – review & editing, Investigation, Formal analysis. Constantinidis Christos: Writing – review & editing, Conceptualization. Luna Beatriz: Writing – review & editing, Supervision, Investigation, Funding acquisition, Conceptualization. Calabro Finnegan: Writing – review & editing, Writing – original draft, Validation, Supervision, Methodology, Investigation, Formal analysis, Data curation, Conceptualization. LeCroy Dylan: Writing – review & editing, Writing – original draft, Validation, Methodology, Formal analysis. Foran Will: Software, Resources, Methodology, Data curation. Sydnor Valerie J.: Writing – review & editing, Visualization, Validation, Methodology.

Declaration of Competing Interest

The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.

The author Beatriz Luna is an Editorial Board Member/Editor-in-Chief/Associate Editor/Guest Editor for Developmental Cognitive Neuroscience and was not involved in the editorial review or the decision to publish this article.

Appendix A. Supplementary material

Supplementary material

Data Availability

Data will be made available on request.