The generalizability of cortical area parcellations across early childhood

Abstract

The cerebral cortex consists of distinct areas that develop through intrinsic embryonic patterning and postnatal experiences. Accurate parcellation of these areas in neuroimaging studies improves statistical power and cross-study comparability. Given significant brain changes in volume, microstructure, and connectivity during early life, we hypothesized that cortical areas in 1- to 3-year-olds would differ markedly from neonates and increasingly resemble adult patterns as development progresses.

Here, we parcellated the cerebral cortex into putative areas using local functional connectivity (FC) gradients in 92 toddlers at 2 years old. We demonstrate high reproducibility of these cortical areas across 1- to 3-year-olds in two independent datasets. The area boundaries in 1- to 3-year-olds were more similar to those in adults than those in neonates. While the age-specific group area parcellation better fits the underlying FC in individuals during the first 3 years, adult area parcellations still have utility in developmental studies, especially in children older than 6 years. Additionally, we provide connectivity-based community assignments of the area parcels, showing fragmented anterior and posterior components based on the strongest connectivity, yet alignment with adult systems when weaker connectivity was included.

Introduction

Understanding the intricate organization of the human brain is a fundamental pursuit in systems neuroscience. Previous research supports the notion that the cerebral cortex is divided into spatially contiguous areas distinguishable by function, architecture, connectivity, and/or topographic organization (Felleman and Van Essen 1991; Eickhoff et al. 2015; Petersen et al. 2024). For example, the differentiation between prestriate and striate areas by a clear histochemical border has been observed in human and monkey fetuses (Kostovic and Rakic 1984). Historically, researchers have utilized histology to find the area organizations (Brodmann 1905; Evans 1992; Carmichael and Price 1994; Amunts et al. 2013; Amunts and Zilles 2015). In recent years, connectivity-based area parcellations, especially those with fMRI data, have become popular as an efficient and non-invasive alternative to parcellate brain areas (Eickhoff et al. 2018, 2015; Glasser et al. 2016; Gordon et al. 2016; Schaefer et al. 2018; Shen et al. 2013). These connectivity-based area parcellations rely on the computation of connectivity strength to other parts of the brain for each voxel/vertex (a.k.a. connectivity profiles) and then group the voxel/vertex into areas with homogeneous connectivity profiles (Eickhoff et al. 2015).

The formation of cortical areas occurs starting from embryonic development. Initially, continuous gradients of signaling molecules within the ventricular zone drive the formation of neurons from their progenitor cells and give rise to a “protomap” (Rakic 1988; Fukuchi-Shimogori and Grove 2001; Bishop et al. 2002; Hamasaki et al. 2004; O’Leary et al. 2007; Stiles and Jernigan 2010; Cadwell et al. 2019). Later, both intrinsic and extrinsic factors refine this “protomap” into discrete areas (O’Leary et al. 2007; Cadwell et al. 2019; Qian et al. 2024). One important contributor to this process is environmental inputs (Greenough et al. 1987; Catalano and Shatz 1998; Tau and Peterson 2010; Smyser et al. 2011), especially those from the thalamocortical axon projections (O’Leary et al. 2007; Vue et al. 2013; Molnár and Kwan 2024). Consequently, early deprivation of sensory inputs could change the functional organization in monkeys (Hubel et al. 1997). The significant increase in exposure to environmental stimuli following birth likely plays an important role in refining of area boundaries shortly after birth. Moreover, synaptic addition and growth of dendrites and spines also enter a logarithmic growth phase in the first few months after birth (Levitt 2003), suggestive of an elevated period of cortical plasticity. Given these factors, cortical areas in neonates likely show low similarity to those in adults (Myers et al. 2024), with increasing resemblance as the brain develops. Furthermore, it has been postulated that developmental changes are not uniform across the brain. The sequence of development has previously been described to follow a sensorimotor-to-association axis (Flechsig 1901; von Economo and Koskinas 1925; Casey et al. 2005; von Economo et al. 2008; Hill et al. 2010; Tau and Peterson 2010; Dean et al. 2015; Smyser and Neil 2015; Smyser et al. 2016; Grayson and Fair 2017; Sydnor et al. 2021), or a posterior-to-anterior axis (Larivière et al. 2020; Li et al. 2024b). Few studies have examined whether the maturation of cortical areas followed either of these patterns.

Many neuroimaging analyses have been conducted at the scale of areas (Arslan et al. 2018; Bijsterbosch et al. 2020; Farahani et al. 2019; Faskowitz et al. 2022; Helwegen et al. 2023; Luppi et al. 2024; Zalesky et al. 2010). Inaccurate area parcellation choice can lead to the mixing of signals (Smith et al. 2011), conceal known community structure (Power et al. 2011), and reduce the prediction accuracy of clinical phenotypes (Abraham et al. 2017). Therefore, choosing an area parcellation scheme that closely reflects the actual area boundaries in the data is of great importance for functional connectivity (FC) analyses (Grayson and Fair 2017).

Neuroimaging analyses often adopt definitions of cortical areas based on adult brains (Shen et al. 2013; Glasser et al. 2016; Gordon et al. 2016; Schaefer et al. 2018). However, the dynamic and rapid development of the brain during infancy (Bethlehem et al. 2022) triggers unique concerns about whether it is valid to apply existing adult area parcellations to early childhood brains (Cusack et al. 2018; Shi et al. 2018; Oishi et al. 2019; Wang et al. 2023). In response, several early childhood area parcellations have been developed in recent years (Scheinost et al. 2016; Shi et al. 2018; Wang et al. 2023; Myers et al. 2024). Despite these advances, having different area parcellations for different age ranges poses a practical challenge for making coherent comparisons in brain organization across development. Thus, many researchers have continued to use adult area parcellations in early childhood studies (Kim et al. 2023; Nielsen et al. 2022; Yates et al. 2023), as well as studies across the lifespan (Betzel et al. 2014; Cao et al. 2014; Zuo et al. 2017; Puxeddu et al. 2020).

When selecting an area parcellation for a specific age group, an important consideration is the extent to which pediatric brain area organization deviates from adult parcellations. However, a systematic examination of parcellations across age groups is lacking. We aim to (i) illustrate how well different area parcellations fit the FC data across individuals at various developmental stages, (ii) quantify the improvement in fit in age-specific parcellations compared to adult parcellations, and (iii) evaluate the potential impact of using an adult parcellation instead of an early childhood parcellation on downstream analyses. If adult parcellations separate the cortical areas with comparable success as early childhood parcellations, utilizing adult parcellation schemes for developmental cohorts would be justifiable. One prior study suggested that this was not the case for neonates (Myers et al. 2024). Here we query whether the adult parcellation would be a reasonable choice for older infants, toddlers, and children.

In the current study, we derive a surface-based area parcellation based on FC local gradient transitions (Cohen et al. 2008; Wig et al. 2014; Gordon et al. 2016) in 92 toddlers at the age of 2 years. To test the reproducibility of our area parcels across groups of subjects and whether the reproducibility followed a uniform distribution across space, we derive parcellations using half the sample (n = 46). To examine differences in FC local gradient transition patterns across development, we quantify the similarity between the boundary maps at different developmental stages. Furthermore, we compare our toddler area parcellation to alternative adult and early childhood parcellations and demonstrate the generalizability and limitations of our area parcellation for application to multiple developmental stages. Finally, we derive the community organization that describes the relationship between the area parcels.

Methods

Experiments were undertaken with the understanding and written consent of each subject or their parents for all datasets used in this project.

Neuroimaging data for deriving area parcellations

One main goal of this paper is to derive an area parcellation for children aged 1 to 3 years. We used two early childhood datasets as the main datasets: eLABE (Early Life Adversity, Biological Embedding) and Baby Connectome Project (BCP) (Table 1). The eLABE study was approved by the Washington University Institutional Review Board (IRB: 201703145). Informed consent was obtained from the parents of all participants. The BCP study was approved by the University of Minnesota and the University of North Carolina Institutional Review Boards, and informed consent was acquired from the parents of all participants. Both datasets were collected with a Siemens Prisma 3 T scanner using HCP-style acquisition parameters. The functional MRI acquisition lasts 420 frames per scan run, with 2–4 runs in the BCP and 2–8 runs in the eLABE 2-year-old data (Y2). Anatomical scan processing and segmentation were conducted using age-specific pipelines (Kaplan et al. 2022). Functional data preprocessing followed established procedures (Power et al. 2014). eLABE (Y2) used the Toddler EPI BOLD preprocessing pipeline, and BCP used the DCAN-Infant v0.0.9 pipeline (Glasser et al. 2013; Donahue et al. 2016; Autio et al. 2020). Motion correction was performed with rigid-body transforms. The functional data were corrected for asynchronous slice time shifts and systematic odd-even slice intensity differences attributable to interleaved acquisition (Power et al. 2012). The data were intensity normalized to achieve a consistent whole-brain mode value, and subsequently resampled to atlas space before being projected onto the 32k_fs_LR standard surface (Van Essen et al. 2012). Subsequently, denoising was accomplished by nuisance regression, with regressors consisting of a 24-parameter Volterra expansion of motion time series, the mean signal over gray-ordinates, and the mean signals derived from white matter and cerebrospinal fluid compartments. The data were bandpass filtered to retain BOLD-specific frequencies and geodesically smoothed with Connectome Workbench (Marcus et al. 2011; Glasser et al. 2013). Frame censoring was performed based on the frame displacement time series (FD > 0.2 mm) following age-specific notch-filtering to exclude respiratory frequencies (Kaplan et al. 2022). Structural and functional scans were manually inspected, and runs/sessions that failed quality controls were discarded. Additionally, participants who were born preterm (<37 weeks' gestational age), had any neonatal ICU experience, or had signs of injury on MRI were also excluded from the analysis. Functional data with fewer than 600 low-motion frames were also excluded. For additional dataset-specific details, see Supplementary Table 1.

Table 1.

Subject demographics for the two infant/toddler datasets. For continuous variables, the mean along with standard deviations in brackets are provided. The group identity was defined as the median age rounded to the nearest whole number.

Group	Age (months)	Age Range (months)	Number of participants	Average retained FD (mm)	Frame retention rate (%)	Acquisition time	% White	% Male
eLABE (Y2)
25 mo	25.2 (1.8)	22–31	92	0.068 (0.021)	92 (9)	20.0 (4.6)	21	59
BCP
10 mo	9.70 (0.71)	8–10	30	0.080 (0.014)	80 (6)	16.4 (5.3)	70	50
12 mo	12.32 (0.80)	11–13	37	0.076 (0.017)	83 (6)	13.3 (4.7)	78	57
16 mo	15.47 (0.88)	14–16	39	0.079 (0.015)	80 (8)	17.6 (5.5)	79	38
19 mo	19.14 (1.52)	17–22	37	0.078 (0.017)	82 (8)	16.4 (5.8)	81	57
25 mo	25.38 (1.60)	23–28	34	0.078 (0.016)	83 (5)	16.7 (5.5)	74	53

A summary of the demographics and image quality of the developmental cohort discovery and validation datasets is provided in Table 1. The cross-sectional and longitudinal age distributions are displayed in Supplementary Fig. 1.

Neuroimaging data for comparing FC boundaries across the lifespan

Additionally, we included the FC boundaries from a young adult dataset (Washington University 120, WU120) used in a widely adopted adult parcellation (Gordon et al. 2016) to act as an adult control. Moreover, we also included the FC boundaries from 131 neonates in the same dataset (eLABE) previously used in a neonatal parcellation (Myers et al. 2024). Acquisition and processing of these datasets followed similar pipelines to the early childhood datasets above, as briefly described below. For dataset-specific details, please refer to Supplementary Table 2.

WU120

Data were collected from 120 healthy young adult participants recruited from the Washington University community during relaxed eyes-open fixation (50% male, ages 19–32). Scanning was conducted using a Siemens TRIO 3 T scanner and included the collection of high-resolution T1-weighted and T2-weighted images, as well as an average of 14 min of resting-state fMRI. Detailed acquisition and processing have been reported previously (Power et al. 2014). The study was approved by the Washington University School of Medicine Human Studies Committee and Institutional Review Board.

eLABE (birth)

The inclusion criteria were the same as those for the eLABE (Y2) cohort. Neuroimaging data were collected in 261 full-term, healthy neonate offspring shortly after birth (the average postmenstrual age of included participants 41.7 weeks, the range 39–45 weeks, and 54% male). 131 participants with the most data following frame censoring were used to create the FC boundaries. Additional details are in Supplementary Table 2.

Neuroimaging data for testing the generalizability of areas across the lifespan

To test for how well our toddler area parcellation and alternative area parcellations generalize to other age groups, we evaluated their cluster validity performance in resting-state fMRI data from additional subjects at the birth (eLABE (Birth)) and year-3 (eLABE (Y3)) time points of the eLABE dataset, Healthy Brain Network (HBN) children dataset, and HCP young adult (HCP-YA) dataset.

eLABE (birth)

This is the same dataset as above. Because 131 of the participants were involved in creating the Myers-Labonte parcellation (Myers et al. 2024), the other 130 participants not used in the generation of the Myers-Labonte parcellation were used to test the parcellation’s performance to prevent circularity.

eLABE (Y3)

The inclusion criteria were the same as those for the eLABE (Y2) cohort. Neuroimaging data were collected from 132 participants at the age of 3 years. Additional participants were excluded based on the quality of structural and functional data and having less than 8 min (600 frames) of low-motion (respiratory-filtered FD < 0.2) data retained, leaving 65 participants (range = 2.93–3.97 years, mean = 3.22 years, SD = 0.32 years, 63% male). The acquisition protocol and processing pipeline were the same as eLABE (Y2).

HBN

Resting-state fMRI data from 493 participants from the first nine releases of the HBN were divided into 10 groups by year (6–15 years). The HBN study is a large, multi-site study of children and young adults aged 5–21 years, all collected in the New York area. Recruitment, consent, and study procedures are described in the data publication (Alexander et al. 2017) and the project website. The study was approved by the Chesapeake Institutional Review Board (https://www.chesapeakeirb.com/). Before conducting the research, written informed consent was obtained from participants aged 18 or older. For participants younger than 18, written consent was obtained from their legal guardians and written assent obtained from the participant. We included the data from two sites (CitiGroup Cornell Brain Imaging Center and Rutgers University Brain Imaging Center).

Data were pre-processed using the Human Connectome Project minimal processing pipeline (Glasser et al. 2013). Additional processing steps (demeaning, detrending, nuisance regression (with regressors consist of a 24-parameter Volterra expansion of motion time series, the mean signal over gray-ordinates), bandpass filtering at 0.008–0.1 Hz to retain BOLD-relevant frequency and frame censoring at respiratory-filtered FD > 0.2 mm were carried out using custom-written Python (v3.8) scripts using the numpy v1.24.4, scipy v1.10.1, nibabel v5.1.0, and pandas v2.0.3 libraries. Each scan session takes 10 min, and all included sessions comprise at least 8 min (600 frames) of data retained following frame censoring. The data was geodesically smoothed to achieve an effective smoothing of a 2.55 sigma Gaussian kernel.

Hcp-YA

Subject recruitment procedures and informed consent forms, including consent to share de-identified data, were approved by the Washington University institutional review board. Resting-state fMRI data from a subset of randomly chosen 40 participants not used to create the Glasser parcellation (Glasser et al. 2016) were selected from the HCP-YA dataset for external validation of the adult dataset to minimize circularity. Data was processed with the same standard preprocessing pipeline as WU 120, except that a low-pass-filtered FD < 0.04 mm was used for frame censoring.

Creation of FC-transition boundary maps and area parcels

We segmented the cortical surface into discrete parcels representing putative cortical areas based on the FC local gradient (Gordon et al. 2016). The FC from each vertex to every other vertex was calculated as Pearson’s correlation of the time series in individual sessions (Supplementary Fig. 2A). For each session, the Fisher-Z-transformed FC from each vertex was correlated with a randomly subsampled set of 594 vertices (1% of the total vertices) to generate an “FC similarity” matrix, which indexed the similarity in FC patterns across vertices (Supplementary Fig. 2B). We used 1% of the vertices for computational efficiency without compromising accuracy (Supplementary Materials). After that, the workbench command “cifti-gradient” was used to calculate the gradient of FC-transition in individual subjects’ surfaces in the standard 32k_fs_LR mesh. The gradient maps were then averaged across all subjects and smoothed with a Gaussian kernel of 2.55 sigma (Supplementary Fig. 2C). A “watershed by flooding” algorithm (Beucher and Meyer 1993) was used to create discrete areas separated by boundaries based on the gradient transitions (Supplementary Fig. 2D). The gradient-based boundary map technique rests on the assumption that FC within a cortical area is relatively uniform and distinct from FC of an adjacent area (Wig et al. 2014), similar to how areas were distinguished by connectivity in macaque monkeys (Felleman and Van Essen 1991). Finally, the boundaries from different gradient maps were averaged to obtain a boundary map that indexed the probability of a vertex being an area boundary (values range between 0 and 1) (Supplementary Fig. 2E).

Discrete parcels (Supplementary Fig. 2F) from a boundary map were created by locating the minima in the boundary map, growing parcels from minima using the watershed algorithm, and merging the watersheds if the median values of boundaries between them are below a threshold (merging threshold, defined as a percentile of the boundary map values). Neighboring parcels with sizes smaller than 15 vertices were merged. Parcels joined only by a single vertex were split. Isolated parcels smaller than 10 vertices were removed. Vertices above 90% in the boundary map values (height threshold) were left as parcel boundaries. The resolution of the parcels depends on the merging threshold, with higher merging thresholds leading to a small number of large parcels and lower merging thresholds leading to a large number of small parcels (as demonstrated in Supplementary Fig. 3).

Parcel reproducibility

To assess the reproducibility of the area parcellation across participants, we repeated the above procedure with non-overlapping split halves of participants in the eLABE (Y2) data 20 times. This results in 20 pairs of boundary maps (Supplementary 2E), and discrete area parcellations (Supplementary 2F) at multiple merging thresholds. For each pair of area parcellations at a given merging threshold, we quantified the overlap in the parcels and in the boundaries from two split halves (See Section: Parcel Similarity Measures). In addition, we divided the brain into 10 equal bins based on either the position along the sensorimotor-association axis (Sydnor et al. 2021) or the posterior–anterior axis and calculated the spatial distribution of parcel reproducibility (Supplementary Fig. 4).

Parcel similarity measures

To evaluate the similarity between cortical parcellations (e.g. generated from different split-halves of the data), we primarily used the Adjusted Rand Index (ARI), focusing on non-boundary vertices to avoid ambiguity. In addition, and for comparability with prior studies, we calculated the Dice coefficient using two approaches: first, by averaging the Dice values across matched parcel pairs, where pairs were defined by maximum overlap (Shen et al. 2013); and second, by applying the Dice coefficient to binarized parcel identity maps, where parcel interiors were assigned a value of 1 and boundaries a value of 0 (Myers et al. 2024). However, this latter approach can be biased by the proportion of the brain covered by parcels—parcellations with wider boundaries may appear artificially more similar due to reduced overlap.

Alternatively, we could assess how similar the parcel boundaries are. Prior study used a dice coefficient (Gordon et al. 2016), but a small spatial shift would drastically reduce the dice coefficient. Therefore, we primarily used the Hausdorff distance, which measures the maximum geodesic distance one needs to travel between two contours in the Conte69 surface atlas (Van Essen et al. 2012). Because it is robust to minor spatial shifts and does not require exact overlap, it serves as a reliable spatial metric. We used two variants of the Hausdorff distance measure: 95% Hausdorff distance (HD95), the 95% Hausdorff distance (HD95), the maximum of the 95^th percentile of the distances () between any point in contour X to the closest point in another contour Y and vice versa (Huttenlocher et al. 1993) and average Hausdorff distance (AHD) defined as the maximum of the mean distance between points on contour X and the closest point on contour Y and vice versa (Müller et al. 2022).We established statistical significance against a spatial null model by inflating the cortical surface to a sphere and randomly rotating the boundaries from one half of the dataset along the x, y, and z axes before computing similarity metrics.

(1)

(2)

Boundary map consistency across age

We applied the same parcellation method (Supplementary Fig. 2) to generate boundary maps from neonates and adults and compared them to the boundary maps derived from the eLABE (Y2) dataset. Similar to compairing binarized parcel boundaries across split halves described above, we thresholded and binarized the boundary maps at top percentile values, and calculated the Hausdorff distance measures.

Evaluation of cluster validity of area parcellation

To evaluate the cluster validity of the area parcellations (i.e. how well they fit the FC data) on each participant’s FC, we used an unbiased cluster validity metric named distance-controlled boundary coefficient (DCBC) for the comparison of area parcellations across different spatial resolutions (Zhi et al. 2022). The similarity (Pearson’s r, with a value between −1 and 1) between FC profiles from all pairs of vertices were calculated, and a difference in within-parcel and between-parcel similarity were calculated for pairs of vertices within each 1 mm geodesic distance bins on the Conte 69 atlas (e.g. between 10 mm to 11 mm). As demonstrated in a prior publication (Zhi et al. 2022), this metric accounts for the spatial smoothness in the data and is relatively unbiased when comparing area parcellations across multiple spatial resolutions (a.k.a. number of parcels). The expected value of DCBC for a random parcellation was zero regardless of the resolution of the parcellation, and a positive DCBC indicates better than random. Thus, no simulation with random null parcellations is necessary to establish a baseline measurement, as opposed to measures like homogeneity Z-score compared to a spatially permuted null (Gordon et al. 2016). As a negative control, we also evaluated an area parcellation that randomly partitioned the brain into 304 equally sized fragments (Icosahedron). Please refer to the Supplementary Materials for implementation details and a comparison with alternative measures.

To further contextualize our results, we computed the same metric on existing area parcellations created using adult or early childhood data. Where necessary, we transformed the area parcellations into the common 32k_fs_LR standard mesh (details in the Supplementary Materials). Table 2 summarizes the area parcellations tested, including the number of parcels, sources, and their original space.

Table 2.

Adult and early childhood area parcellations.

Name	Number of parcels	Citation	Original Space
Gordon	333	Gordon et al. 2016	32k_fs_LR
Glasser	360	Glasser et al. 2016	32k_fs_LR
Schaefer	400	Schaefer et al. 2018	32k_fs_LR
AAL	82	Tzourio-Mazoyer et al. 2002	MNI152
Desikan	70	Desikan et al. 2006	32k_fs_LR
Shen	200	Shen et al. 2013	MNI152
Myers-Labonte	283	Myers et al. 2024	32k_fs_LR
Tu	326	Current study	32k_fs_LR
Wang	864	Wang et al. 2023	32k_fs_LR
Scheinost	87	Scheinost et al. 2016	MNI152
Shi 1 Yr	194	Shi et al. 2018	Age-specific T1 (Shi et al. 2011)
Shi 2 Yr	205	Shi et al. 2018	Age-specific T1 (Shi et al. 2011)
Icosahedron (control)	304	Zhi et al. 2022	32k_fs_LR

Comparing the toddler area parcellation to narrower-age early childhood area parcellations

In addition to the toddler area parcellation from the eLABE (Y2) data, we also generated age-specific area parcellations with five narrower age groups of the BCP participants using a merging threshold of 65%. To test whether finer age-specific area parcellations improve cluster validity for the corresponding age group, we calculated the DCBC for these five narrower age group parcellations on an additional subset of validation sessions in the same age range (n = 73 sessions from 51 participants, age range 8–29 months) divided into the same age bins. These validation sessions included a more recent release of the BCP data collected at both the University of Minnesota and the University of North Carolina Chapel Hill sites with the same acquisition and largely the same processing using the DCAN-Infant pipeline v0.0.22, where zero-padding at both ends of the BOLD time series was introduced prior to band-pass filtering to minimize the distortions in the edges of the time series.

Practical implications of using early childhood and adult parcellations

A good area parcellation should produce functional connectomes (a.k.a. pairwise FC between all pairs of areas) that yield satisfactory behavioral phenotype prediction accuracy (Kong et al. 2023) and demonstrate decent test–retest reliability (Tozzi et al. 2020). We thus compare the prediction accuracy of age using FC from the BCP dataset based on the present 2-year-old parcellation (Tu (326)) and the Gordon parcellation (Gordon et al. 2016), the best-performing early childhood and adult area parcellations based on cluster validity, respectively. We constructed a functional connectome with the first 7.2 min (600 frames for sessions with TR = 0.8 and 540 frames for sessions with TR = 0.72) of low-motion fMRI data in each subject in the BCP dataset using the area parcellations and applied a linear support vector regression for the prediction of age (Li et al. 2024a). In addition, we assessed the test–retest reliability of individual edges in the parcellated FC using an intraclass correlation coefficient measure (ICC (3,1)) (Shrout and Fleiss 1979; Tozzi et al. 2020) in the first 5 minute of low-motion data in non-overlapping runs. Details are provided in the Supplementary Materials.

Identification of community structure in 2-year-olds

To characterize the relationship between the area parcels, we identified the community structure with the Infomap algorithm on the graph representation using area parcels as nodes and the FC between parcels as edges following prior research (Rosvall and Bergstrom 2008; Gordon et al. 2016). For each participant in the eLABE (Y2) dataset (n = 92), we created a parcellated time series by calculating the mean within-parcel time series over each of the parcels from the dense grayordinate time series in 32k_fs_LR space with the workbench command “wb_command cifti-parcellate.” We then cross-correlated these parcellated time series to generate a parcel-wise correlation matrix. Parcel-to-parcel correlation matrices were Fisher z-transformed and averaged across all participants to obtain a group-average FC matrix. Next, to reduce the impact of non-neuronal sources of inflation in short-distance correlation (e.g. data processing, subject motion), we applied an exclusion distance of 30 mm on the correlation matrix. After thresholding, a weighted sparse graph (edge density in steps of 0.25% ranging from 0.25% to 20%) was entered as the input to the Infomap algorithm. Lastly, a consensus was obtained from manual examination of the communities at different thresholds to identify reliable networks across multiple thresholds that match with the expected network topography based on prior knowledge. In addition, we also examined the networks at relatively lower edge density thresholds, keeping the naming convention and colors similar to what was described in an earlier publication (Myers et al. 2024).

Results

Area parcellation in 2-year-olds is reproducible across participants

The reproducibility of the area parcellations across participants was evaluated using split-half sampling 20 times (Fig. 1A-B). We found that the reproducibility was highest around a merging threshold of 60–80%, significantly larger than the spatially permuted null model (Fig. 1C-E). Based on manual inspection of the boundary map and the commonly reported granularity of area parcels in adult area parcellations (Glasser et al. 2016; Gordon et al. 2016), we settled on a merging threshold of 65% for our main parcellation, which produced 324–391 parcels across 20 split-haves (Supplementary Fig. 5A). For the remaining sections, the final area parcellation using all data in eLABE (Y2) (n = 92) and a merging threshold of 65% was used for evaluation, hereafter referred to as “Tu (326).” At the merging threshold of 65%, ARI = 0.66 ± 0.02, Z-score compared to the null model = 14.3, parcel averaged dice coefficient = 0.62 ± 0.01, Z-score compared to the null model = 15.2. The dice coefficient for the binarized parcel map is 0.87 ± 0.002, Z-score compared to the null model = 8.46. Similar results were obtained when comparing the reliability of binarized boundary maps (see Supplementary Materials). Furthermore, we found that the sensorimotor areas tend to have higher reproducibility than the association areas (Fig. 1F-H) and that the posterior areas tend to have higher parcel reproducibility than the anterior areas (Fig. 1I-K) when the same parcel reproducibility metrics were computed for 10 equal divisions along the sensorimotor-association axis (Supplementary Fig. 4A) and the posterior–anterior axis (Supplementary Fig. 4B).

Fig. 1.

Parcel reproducibility between split halves. A) Area parcels and area boundaries from an example first split-half and second split-half. B) The overlap in black between the parcels (left) and boundaries (right). C) Adjusted Rand index (ARI). D) Parcel-average dice coefficient. E) Dice coefficient on binarized parcels. The blue line and shaded area show the actual values and the standard deviation across 20 splits. The black line and shaded area illustrate the mean and 95% confidence interval of the spatially permuted null from one example split. The dashed line shows merging threshold = 65%. F-H: The same metrics as C-E but separated into 10 bins along the sensorimotor-association axis at merging threshold = 65%. I-K: The same metrics as C-E but separated into 10 bins along the posterior–anterior axis at merging threshold =65%. The colors in the individual data point in F-K match the bin colors in Supplementary Fig. 4.

Functional connectivity transition boundaries in 2-year-olds are consistent within the group and resemble those in adults more than those in neonates

We compared the FC boundary maps from the 2-year-olds (Fig. 2A-B) to those generated from adults (Fig. 2C) and neonates (< 2 weeks, Fig. 2D) by first binarizing the vertices at the top percentile of boundary probabilities (using a threshold of 15–55%) (Supplementary Fig. 6). FC boundaries in 2-year-olds were spatially closer to adult boundaries (e.g. HD95 = 7.61 ± 0.24 mm, AHD = 2.22 ± 0.03 mm for the top 35% vertices) compared to neonate boundaries (e.g. HD95 = 8.68 ± 0.01 mm, AHD = 2.63 ± 0.01 mm for the top 35% vertices) (Fig. 2E-F). Across the five early childhood age bins (median age 10, 12, 16, 19, 25 months, Table 1) in the BCP dataset, FC boundaries were considerably similar, despite a small trend of elevated similarity across groups with a smaller age difference (HD95 5 mm for the top 35% vertices, Supplementary Fig. 7).

Fig. 2.

Similarity of boundary maps across age groups. A) The FC boundary map in an example first split half. B) The FC boundary map in an example second split half. C) The FC boundary map in an adult dataset (WU 120). D) The FC boundary map in a neonate dataset (eLABE (birth)). (E) 95% Hausdorff distance (HD95) indexes the spatial similarity of the boundaries between eLABE (Y2) Split-I and those from eLABE (Y2) Split-II, adult, and neonate. The shaded area indexes the 95% confidence interval for the HD95 between the FC boundary in eLABE (Y2) Split-I and 1000 spatially permuted nulls of eLABE (Y2) Split-II. F) Same as E but using AHD. Lower HD95 and AHD indicate more similar boundaries.

Local gradient-based 2-year-old area parcellation provides the best cluster validity for children at 1–3 years

Using FC profiles from the eLABE (Y2) individuals, we evaluated the cluster validity of the local gradient-based 2-year-old parcellation against several existing adult and early childhood area parcellations (Fig. 3A), as well as a regular hexagonal parcellation of a sphere (Icosahedron) with 304 parcels (Supplementary Fig. 8). We observed substantial variation in cluster validity within both adult and early childhood parcellation groups. Among adult parcellations, the Gordon parcellation performed best. In contrast, the Tu (326) parcellation performed best among early childhood parcellations (Fig. 3B). Notably, all examined adult and early childhood parcellations examined—except for AAL (82) and Desikan (70)—showed DCBC values significantly greater than 0 (FDR-corrected P < 0.05). In contrast, DCBC for the control Icosahedron (304) parcellation was not significantly above 0. A repeated measures ANOVA with the 13 parcellations as the within-subject factor was conducted on the 13x92 DCBC matrix, revealing significant differences in DCBC across parcellations (F(12,1092) = 508.64, P < 0.001). Post-hoc paired t-tests indicated that Tu (326) exhibited superior cluster validity (Cohen’s d > 2.0) compared to alternative adult and early childhood parcellations in eLABE (Y2) individuals (Supplementary Fig. 9A).

Fig. 3.

Cluster validity for different area parcellations evaluated with a distance-controlled boundary coefficient (DCBC) measure. A) Adult area parcellations and early childhood area parcellations. B) DCBC quantified in individuals in the same eLABE (Y2) dataset used to derive the Tu (326) parcels. C) DCBC quantified in individuals in an independent dataset (BCP). ^*P < 0.05 after FDR correction for one-sample t-test against 0. As a convention, we noted the number of parcels of a particular parcellation scheme in parentheses, e.g. Gordon (333) means Gordon parcellation with 333 parcels.

One caveat is that this evaluation was performed on the same dataset used to generate the parcels. To address this, we further validated the parcellations using an independent dataset (BCP) (Fig. 3C). The Gordon (333) and Tu (326) parcellations remained the top performers within the adult and early childhood parcellations, respectively, confirming the robustness of our results. Another repeated-measures ANOVA (13 parcellations as the within-subject factor) confirmed significant differences in DCBC (F(12,396) = 100.92, P < 0.001) but no interaction between age bins and parcellations (F(48,2064) = 0.76, P = 0.88). Post-hoc paired t-test indicated that Tu (326) exhibited superior cluster validity (Cohen’s d > 1.2) at 8–30 months (Supplementary Fig. 9B). Additionally, we computed DCBC for all five BCP age groups (Supplementary Fig. 10). A repeated measures ANOVA (13 parcellations as the within-subject factor, five age bins as between-subject factor) on the 13 × 177 DCBC matrix revealed significant DCBC difference across parcellations, F(12,2064) = 551.31, P < 0.001), but no interaction between age bins and parcellations, F(48,2064) = 0.76, P = 0.88). Similar results were observed when calculating homogeneity Z-scores at the group-average level (Supplementary Figs. 11–12). Further are provided in the Supplementary Materials.

Age-specific early childhood area parcellations show comparable cluster validity to the 2-year-old parcellation

We further generated age-specific area parcellations using the BCP dataset, applying a 65% merging threshold across five narrower age windows (ranging from 352 to 380 parcels; Fig. 4A). These parcellations demonstrated moderate similarity to one another, with adjusted Rand index (ARI) values between 0.5 and 0.6. To assess their performance, we computed the DCBC for the age-specific parcellations, the Tu (326) and the Gordon (333) using additional sessions of BCP data from a more recent release. Repeated-measures ANOVAs (with the three parcellations as the within-subject factor) revealed significant differences across parcellations for all age groups: 10 months (F(2, 18) = 15.86, P < 0.001), 12 months (F(2, 20) = 21.52, P < 0.001), 16 months (F(2, 24) = 21.74, P < 0.001), 19 months (F(2, 36) = 49.01, P < 0.001), and 25 months (F(2, 38) = 48.61, P < 0.001). Post-hoc two-tailed paired t-tests indicated that the age-specific parcellation outperformed the Tu (326) parcellation only at 10 months (FDR-corrected P = 0.038; Fig. 4B) but performed significantly worse at 19 months (FDR-corrected P = 0.0065). Since the Tu (326) parcellation was derived from a separate dataset, these results support its generalizability and utility across the 1- to 2-year age range.

Fig. 4.

Cluster validity for age-specific early childhood area parcellations. A) Age-specific early childhood area parcellations in narrower age windows. B) DCBC on a secondary validation dataset of held-out BCP participants using the age-specific, Tu (326), and Gordon (333) parcellations. ^**P < 0.01, ^***P < 0.001. FDR-corrected for 3 paired t-tests.

We also examined whether age-specific parcellations (Wang et al. 2023)—derived from 3-, 6-, 9-, 12-, 18-, and 24-month groups—better fit individual FC profiles at corresponding ages. However, we found no clear evidence that age-matched parcellations provided superior fits. All Wang parcellations exhibited low DCBC values (<0.02; Supplementary Fig. 13), suggesting minimal advantage over the Tu (326) parcellation.

Adult parcellations based on FC exhibit higher cluster validity from age 6 onward

To assess the lifespan trajectory of cluster validity, we evaluated adult and early childhood area parcellations using FC data from individual neonates (eLABE (Birth)), 3-year-olds (eLABE (Y3)), school-aged children (HBN), and young adults (HCP-YA). Neonate data showed optimal fit with the Myers-Labonte (283) parcellation (Fig. 5A), whereas 3-year-old data were best characterized by the Tu (326) parcellation (Fig. 5B). These findings remained consistent when using an alternative of the Myers-Labonte parcellation with broader cortical coverage (Myers et al. 2024;Supplementary Fig. 15). For older age groups. The Gordon (333) parcellation demonstrated superior fit: it best captured FC patterns in both school-aged children (Fig. 5C-E, Supplementary Fig. 14) and young adults (Fig. 5F). Comprehensive pairwise comparisons across parcellations for each age group are provided in Supplementary Fig. 16.

Fig. 5.

Cluster validity for different adult and early childhood parcellations across other developmental stages. A) a neonate dataset eLABE (birth), B) an older toddler dataset eLABE (Y3), C–E) a children dataset HBN, and D) a young adult dataset HCP. ^*P < 0.05 after FDR-correction.

Practical implications of early childhood versus adult parcellations

We conducted two representative analyses to evaluate the practical differences between early childhood and adult parcellations: (i) predicting chronological age from parcellated functional connectomes and (ii) assessing test–retest reliability. For age prediction, accuracy increased with parcel number but plateaued at ~300 parcels (Fig. 6A), with no systematic advantage for either adult or early childhood parcellations. The spatial distributions of edges with top 5% positive correlations with age and top 5% negative correlations with age were remarkably consistent between the optimal early childhood (Tu (326)) and adult (Gordon (333)) parcellations (Fig. 6B-C). Medial visual, motor, and medial parietal areas contained the highest density of age-correlated edges, whereas lateral frontal areas contained the lowest density of age-correlated edges (Fig. 6D-E). This pattern remained stable across other parcellations (Supplementary Figs. 17–19), though coarse anatomical parcellations (e.g. Desikan (70)) occasionally underrepresented certain regions (e.g. motor cortex). Regarding test–retest reliability, both the Tu (326) and Gordon (333) parcellations revealed lower intraclass correlation coefficients (ICC) in the motor areas and lateral/medial prefrontal cortices (Fig. 7). This reliability pattern was consistent across other parcellations (Supplementary Figs. 20–21).

Fig. 6.

Connectome-based phenotype prediction using adult and early childhood parcellations. A) Multivariate prediction accuracy of age at scan. Error bars show the mean and standard deviation of correlation of actual age and predicted age at scan across 1000 random samples. The red line shows a logarithmic fit to the data. The bolded symbol shows the best adult (Gordon (333)) and best early childhood (Tu (326)) parcellations in Fig. 3. B) Top 5% positive edges in age-FC correlation magnitude (top) and top 5% negative edges in age-FC correlation magnitude (bottom), nodes represent parcel centroids in Tu (326) C) same as B but in Gordon (333). D) Number of significant age-FC correlation edges from each parcel in Tu (326). E) Number of significant age-FC correlation edges from each parcel in Gordon (333).

Fig. 7.

Test–retest reliability from the connectomes using adult and early childhood parcellations. A) Calculation of test–retest reliability of edges. AP = anterior-to-posterior, PA = posterior-to-anterior; example FC matrix was sorted by the Gordon et al. 2016 network orders. B) The edges with a “good” reliability (ICC = 0.60–0.75) using Tu (326). C) The edges with a “good” reliability (ICC = 0.60–0.75) using Gordon (333). D) Mean edge test–retest reliability for edges connected to areas in the brain using Tu (326). Nodes represent parcel centroids. E) Mean edge test–retest reliability for edges connected to areas in the brain using Gordon (333).

Functional network organization from area parcels

Cortical areas interact differentially to form large-scale functional networks (Power et al. 2011; Yeo et al. 2011). Using the Tu (326) parcels as graph nodes, we derived data-driven functional network assignments optimized for reliability across edge densities (Supplementary Video). At higher edge densities in 2-year-olds, we observed cohesive integration of the default and frontoparietal networks, unlike the fragmented anterior–posterior organization seen in neonates (Myers et al. 2024; Sylvester et al. 2022; Fig. 8A). This suggests strengthening of long-range FC between ages 0–2 years. At lower edge densities, networks showed more localized organization. The default mode network is segregated into four components: posterior default, inferior frontal gyrus, dorsomedial PFC, and ventromedial PFC. This localized pattern differs markedly from the distributed default mode network sub-components seen in adults (Andrews-Hanna et al. 2010; Yeo et al. 2011; Gordon et al. 2020; Lynch et al. 2024). Similarly, the frontoparietal network is separated into posterior frontoparietal, lateral PFC, and anterior PFC (Fig. 8B). The visual network differentiated into central, peripheral, and streams subsystems, with the latter sometimes reported as a sub-component of the dorsal attention network (Yeo et al. 2011; Du et al. 2024). Seed-based connectivity maps further illustrated the progression of long-range FC strength across development (Supplementary Fig. 22).

Fig. 8.

Assigned community identities for each parcel. A) Consensus community assignment for 12 networks. B) Finer division into 19 networks. Acronyms: PFC = prefrontal cortex, SM = Somatomotor.

Discussion

We developed a FC-based cortical area parcellation specifically optimized for 2-year-olds. Compared to existing adult and early childhood parcellations, our approach demonstrated superior fit for children aged 1–3 years old across two independent datasets, establishing its utility for future neuroimaging studies in this age range. Notably, the best-performing adult parcellations also provided good fit for 1–3 year-olds—comparable to some existing early childhood area parcellations—and outperformed our 2-year-old parcellation in children 6 years and older. Combined with the results that area boundaries in 2-year-olds are more similar to those in adults than those in neonates, using adult area parcellations in developmental studies represents a reasonable approach, particularly for children beyond infancy. These findings support the hypothesis that cortical areas undergo substantial refinement during the first postnatal year before stabilizing into more adult-like configurations. Our work advances understanding of cortical arealization in early human development while providing practical guidelines for neuroimaging studies of developmental cohorts. The parcellation scheme we developed offers a tailored tool for investigating FC patterns during this crucial transitional period of brain development.

Variability across the sensorimotor-association hierarchy

We observed greater consistency in area boundaries at the sensorimotor end compared to the association end of the sensorimotor-association hierarchy. This pattern may reflect two complementary factors: (i) lower interindividual variability in sensorimotor systems (von Economo and Koskinas 1925; von Economo et al. 2008; Mueller et al. 2013; Gratton et al. 2018; Kong et al. 2019; Li et al. 2019; Sydnor et al. 2021), and (ii) sharper cytoarchitectonic boundaries in these regions, as documented in macaque neuroanatomy (Lewis and Van Essen 2000). These findings align with extensive evidence demonstrating earlier maturation of sensorimotor cortices relative to association cortices, including FC (Smyser et al. 2011; Gao et al. 2015a), cortical thickness (Ahmad et al. 2023), surface expansion (Hill et al. 2010; Garcia et al. 2018), gray matter density (Gogtay et al. 2004), intrinsic activity (Sydnor et al. 2023, 2021), and intrinsic time scales (Truzzi and Cusack 2023). The earlier developmental maturation of sensorimotor cortices may constrain their plasticity relative to association cortices (Hill et al. 2010). This reduced plasticity could render sensorimotor cortices less susceptible to environmental influences (Gao et al. 2017), potentially explaining the observed stability in their area boundaries across individuals.

Developmental trajectory of FC boundaries

Current models of cortical arealization propose a dual mechanism involving: (i) genetically-driven morphogen gradients establishing initial patterning, and (ii) the activity-dependent refinement of boundaries (Cadwell et al. 2019) through thalamocortical inputs (O’Leary et al. 2007). Our findings align with this framework by demonstrating that FC transition boundaries in 2-year-olds more closely resemble those in adults than in neonates. This suggests rapid boundary refinement during early infancy, with developmental rates slowing substantially by 1–2 years of age. Complementary evidence from Wang and colleagues shows relatively stable FC-gradient transition boundaries between 3 and 24 months (Wang et al. 2023). This timeline corresponds with the most rapid cortical surface expansion period, which begins to decelerate after 3 months (Bethlehem et al. 2022; Huang et al. 2022).

Notably, while 2-year-old boundaries show greater similarity to adult patterns, they still maintain significant organization-level resemblance to neonatal boundaries compared to spatially permuted null models. This residual similarity supports the existence of an intrinsic proto-architecture present at birth, consistent with embryonic morphogen-driven patterning (O’Leary et al. 2007; Tau and Peterson 2010; Smyser et al. 2011; Cadwell et al. 2019).

Optimization of parcellation resolution for biological validity and utility

Our parcel resolution of 326 parcels aligns with existing adult and early childhood area parcellations and corresponds to current estimates of 300–400 cortical areas in humans based on multimodal evidence (Van Essen et al. 2012; Glasser et al. 2016; Van Essen and Glasser 2018). This choice enhances biological interpretability while addressing several practical considerations. First, automated approaches are inherently sensitive to data variations and parameter choices, often yielding multiple valid solutions (Supplementary Fig. 3; Van Essen and Glasser 2018). Second, resolution significantly impacts global graph metrics (Zalesky et al. 2010; Arslan et al. 2018), making our comparable resolution particularly valuable for cross-age comparisons between children and adults.

The absence of 3D histological atlases for early cortical maturation precludes direct histological validation. While a fine-grained early childhood area parcellation exists (Wang et al. 2023), our analyses indicate limited generalizability across processing pipelines and datasets. Importantly, multiple lines of evidence suggest that: (i) demographic and behavioral prediction from FC data plateaus at ~ 300 parcels (Arslan et al. 2018; Kong et al. 2023), and (ii) the correspondence between FC gradients and the Mesulam hierarchy emerges consistently across parcellation schemes with > 300 nodes (Vos de Wael et al. 2020). These findings suggest minimal practical advantage for finer-grained parcellations in common analyses like brain network graph properties or multivariate phenotype prediction. Recognizing that optimal resolution may vary by application (Zalesky et al. 2010; Schaefer et al. 2018), we have made our 2-year-old area parcellation available at multiple resolutions. However, we caution that higher-resolution versions may show reduced generalizability across individuals and datasets, and should therefore be used with judiciously.

Performance of adult parcellations in developmental cohorts

Our analyses revealed that while the optimal adult parcellation (Gordon (333)) showed poorer fit to FC data in 0–3 year-olds compared to the best early childhood parcellations, it nevertheless demonstrated significant cluster validity (DCBC > 0) and outperformed the regular hexagonal (icosahedron) control parcellation. These findings indicate meaningful correspondence between adult cortical areas and those present from the neonatal period through age 3 years. For extended discussion of these results about prior literature, see Supplementary Materials.

Comparable outcomes using adult versus early childhood parcellations for key analyses

Our results demonstrate that age prediction accuracy from FC data increases with parcel number before plateauing around 300 parcels, consistent with previous reports (Arslan et al. 2018; Kong et al. 2021). While one study reported marginal effects of atlas choice for predicting psychological and clinical traits (Dadi et al. 2019), this may reflect differences in parcel number between data-driven and predefined approaches rather than fundamental advantages of either method. Regarding test–retest reliability, we replicated adult findings showing greater stability in temporal and parietal connections (Tozzi et al. 2020), though overall reliability was lower in early childhood (Fig. 7). These findings align with reported infant edge-level reliability of mean ICC around 0.14–0.37 (Dufford et al. 2021; Wang et al. 2021). Several methodological factors likely contribute to this reduced reliability: (i) limited scan duration (5 min) for test and retest segments, (ii) varying phase-encoding directions, and (iii) potential sleep state transitions during scanning (Mitra et al. 2017; Lee et al. 2020). While our limited analyses support previous studies employing adult parcellations (Kardan et al. 2022; Nielsen et al. 2022), we emphasize that this does not imply adult parcels precisely represent early childhood cortical organization. Rather, these results suggest that adult parcellations may yield qualitatively similar conclusions for certain analyses despite potential neurodevelopmental differences.

Emerging network organization in 2-year-olds shows both adult-like and developmental features

We identified fragmented components of canonical adult functional systems in 2-year-olds, consistent with previous studies using similar approaches in this age range (Eggebrecht et al. 2017; Kardan et al. 2022; Wang et al. 2023). This finding supports the well-established developmental trajectory where long-range FC matures later than short-range FC (Smyser et al. 2011, 2010; Smyser and Neil 2015; Spisák et al. 2014; Sylvester et al. 2022; Thomason et al. 2015). However, when incorporating weaker connections, the network topography in 2-year-olds showed striking similarities to adult functional networks (Power et al. 2011; Yeo et al. 2011; Gordon et al. 2016; Ji et al. 2019). Independent validation demonstrated that these network assignments fit children aged 1–2 years and substantially outperformed the young adult network assignments (Tu et al. 2025).

Key developmental differences emerged in network organization. First, unlike adults, the temporal lobe remained largely segregated from the default mode network. Second, the motor hand/foot network incorporated portions of the inferior parietal lobule and posterior insula, likely driven by the elevated connectivity between inter-effector areas and the action-mode/cingulo-opercular network (Gordon et al. 2023). This potentially reflects developmental plasticity supporting multisensory integration. Third, we observed encroachment of salience network regions into the cingulo-opercular/action mode network.

Methodologically, we note that the Infomap algorithm’s tendency to identify localized clusters when analyzing strong FC (particularly prominent in developmental cohorts due to distance-dependent connectivity) may influence network detection. Alternative approaches that account for distance dependence (Zamani Esfahlani et al. 2020; Sylvester et al. 2022) may better reveal adult-like functional networks (Petersen and Sporns 2015). Rather than binary classification of network connectivity, we emphasize evaluating algorithm performance across edge densities relative to adult topography. Accordingly, we provide two complementary models: a 12-network model resembling adult functional networks, and a 19-network model with granularity similar to neonatal networks (Myers et al. 2024).

We caution that the biological significance of these FC-derived networks as true “functional systems” (Power et al. 2011; Yeo et al. 2011; Wig 2017) remains uncertain. They likely represent immature precursors of adult systems (Gao et al. 2015b, 2015a).The biological validity of the fragmented components requires validation through task-based imaging in early childhood (Yates et al. 2022, 2021) in future research. Researchers should consider these limitations when applying our network models, particularly the need for future validation of the fragmented components we identified.

Practical recommendations for selecting parcellation schemes

Age-specific parcellations offer optimal fidelity for their target age groups but present challenges for cross-age comparisons due to variable node and edge counts. While cortical development might theoretically preclude a universal parcellation, our results demonstrate that the Tu (326) parcellation generalizes well across 1–3 year-olds. We therefore propose two alternative approaches, each with distinct advantages:

For studies prioritizing cross-age consistency, canonical adult parcellations (e.g. Gordon (333)) provide standardized nodes across developmental stages (Oishi et al. 2019). Although this approach may sacrifice age-specific optimization, our findings suggest minimal practical impact on common analyses like connectome-based age prediction. Alternatively, individualized approaches can better capture neurodevelopmental variation. Current methods include: (i) deriving participant-specific parcellations from group priors (Chong et al. 2017; Li et al. 2017, 2019, 2022; Zhao et al. 2020; Kong et al. 2021; Qiu et al. 2022), (ii) latent space embedding for cross-participant alignment (Langs et al. 2016; Haxby et al. 2020; Nenning et al. 2020), and (iii) precision functional mapping of intensively sampled individuals (Laumann et al. 2015; Gordon et al. 2017). These techniques may prove particularly valuable for developmental studies requiring fine-grained individual differences.

Limitations and future directions

While non-human primate studies have established the neurophysiological basis of FC observed in fMRI (Logothetis et al. 2001; Vincent et al. 2007; Pagani et al. 2023) and its relationship to anatomical connectivity (Vincent et al. 2007; Adachi et al. 2012; Pritschet et al. 2020), there are inherent limitations associated with defining area boundaries based solely on fMRI (Eickhoff et al. 2015; Van Essen and Glasser 2018). First, FC estimates from fMRI do not fully account for critical neurobiological factors, including the balance of excitatory and inhibitory signals within the maturing cortex (Ben-Ari 2002; Markram et al. 2004), the retraction of cortical fibers, and the growth or arborization of dendrites (Hua and Smith 2004; Stiles and Jernigan 2010). Moreover, FC gradients reflect functional co-activation patterns rather than strict architectonic boundaries (Van Essen and Glasser 2018). For example, FC gradients segregate motor cortex into approximately foot, hand, and face patches rather than elongated architectonic divisions (Glasser et al. 2016). Whether participants have their eyes closed during the scan also seems to affect visual area FC (Laumann et al. 2015; Van Essen and Glasser 2018). Therefore, as with all FC-based area parcellations, our putative areas may not represent exactly the neurobiological boundaries.

Several methodological considerations merit attention. Our current parcellation focuses exclusively on cortical organization in 1–3-year-olds, and future work should incorporate subcortical and cerebellar structures and consider prenatal differentiation of areas. The group atlas approach can be affected by multiple factors, including acquisition parameters, resolution, inter-individual variability in functional and anatomical organizations (Shen et al. 2013; Ahmad et al. 2023). Future research with smaller voxels or a better T2* protocol to increase signal-to-noise ratio may further improve the quality of the group area parcellation.

Several potential confounds should be noted. Minor differences in acquisition and processing (Supplementary Table 1) could potentially impact the appearance of area boundaries. In addition, future studies should also investigate how much of the differences between neonate and their older-age counterparts could be attributed to the challenges in the registration of the neonate’s brains due to their tissue properties and anatomical differences from the adult brain. Additionally, when testing the generalizability of area parcellations to neonates, 2-year-olds, and 3-year-olds, we used data from overlapping subjects from the eLABE longitudinal dataset, which might have provided a slight advantage to the Myers-Labonte (283) and Tu (326) area parcellations. Sleep-state effects on FC patterns (Mitra et al. 2017) are particularly relevant given developmental changes in sleep architecture (Kahn et al. 1996) and documented FC differences across active and quiet sleep (Lee et al. 2020), and across sleeping and movie-watching (Yates et al. 2023; Tu et al. 2025) in infants and toddlers. Finally while necessary for comparison, the transformation of volumetric parcellation to surface space may have unintentionally favored surface-optimized parcellations.

Future directions should prioritize sleep-state controlled acquisitions (Mueller et al. 2025), multimodal validation against cytoarchitecture and task activation, improved cross-modal registration techniques, and expanded age coverage, including prenatal development. While acknowledging these limitations, FC-based parcellations remain valuable for dimensionality reduction and developmental comparisons when interpreted appropriately (Van Essen and Glasser 2018).

Conclusion

In this study, we developed FC gradient-based cortical parcellations specifically optimized for 2-year-olds, providing a valuable tool for investigating FC patterns in children aged 1–3 years. Our analyses revealed that cortical area boundaries in 2-year-olds demonstrate greater similarity to adult configurations than neonatal patterns. Among all parcellations evaluated, our approach achieved superior cluster validity for the 1–3 year age range across two independent datasets. While adult parcellations showed reduced fit compared to age-appropriate alternatives, they nevertheless captured meaningful organizational principles in early childhood FC data.

These findings provide empirical support for two key conclusions: First, they validate the practice of using adult parcellations in studies of early childhood FC when age-specific alternatives are unavailable. Second, they reinforce the hypothesis that the most substantial refinement of cortical areas occurs primarily during the first postnatal months, with relative stabilization thereafter. Beyond these theoretical contributions, our work offers practical guidance for neuroimaging studies of developing populations, bridging a critical methodological gap in developmental neuroscience research.

Author contributions

Jiaxin Cindy Tu (Conceptualization, Formal analysis, Funding acquisition, Writing—original draft, Data curation, Writing—review & editing), Michael J Myers (Resources, Software, Writing—review & editing), Wei Li (Formal analysis, Writing—review & editing), Jiaqi Li (Methodology, Software, Writing—review & editing), Xintian Wang (Data curation, Writing—review & editing), Donna Dierker (Data curation, Writing—review & editing), Trevor Day (Data curation, Writing—review & editing), Abraham Snyder (Writing—review & editing), Aidan Latham (Data curation, Writing—review & editing), Jeanette K Kenley (Data curation, Writing—review & editing), Chloe M Sobolewski (Data curation, Writing—review & editing), Yu Wang (Data curation, Writing—review & editing), Alyssa K Labonte (Resources, Writing—review & editing), Eric Feczko (Data curation, Supervision, Writing—review & editing), Omid Kardan (Data curation, Writing—review & editing), Lucille A Moore (Data curation, Writing—review & editing), Chad M Sylvester (Supervision, Writing—review & editing), Damien A Fair (Supervision, Writing—review & editing), Jed Elison (Funding acquisition, Supervision, Writing—review & editing), Barbara B Warner (Funding acquisition, Writing—review & editing), Deanna M Barch (Funding acquisition, Writing—review & editing), Cynthia E Rogers (Funding acquisition, Writing—review & editing), Joan L Luby (Funding acquisition, Writing—review & editing), Christopher D Smyser (Funding acquisition, Writing—review & editing), Evan M Gordon (Resources, Software, Writing—review & editing), Timothy Laumann (Resources, Software, Writing—review & editing), Adam Thomas Eggebrecht (Conceptualization, Supervision, Writing—review & editing), Muriah Wheelock (Conceptualization, Funding acquisition, Supervision, Writing—original draft, Writing—review & editing).

Funding

This work is partially supported by the CCSN fellowship from the McDonnell Center for Systems Neuroscience at Washington University School of Medicine in St. Louis to JCT and from NIH grants including NIBIB K99/R00 EB029343 and NICHD R01 HD115540 to MDW, R01MH122389 and R01MH134966 to CMS. The Early Life Adversity, Biological Embedding (eLABE) study was supported by NIMH R01 MH113883. The Baby Connectome Project was supported by NIMH R01 MH104324 and NIMH U01 MH110274.

Conflict of interest statement: The authors declared no competing interests directly related to this manuscript.

Data and code availability

Baby Connectome Project data are available for download at the NIH Data Repository website: https://nda.nih.gov/edit_collection.html?id=2848. Early Life Adversity, Biological Embedding (eLABE) data are available through request at https://eedp.wustl.edu/research/elabe-study/.

All analyses, unless otherwise stated, were implemented with custom MATLAB scripts in the R2020b release. All visualizations were created with custom MATLAB scripts or Connectome Workbench Version 1.5.0.

The code for the generation and evaluation of parcellations is adapted from the MSCcodebase and the DCBC toolbox. Code and parcellations mentioned in the manuscript are available to download here.