Few sex differences in regional gray matter volume growth trajectories across early childhood

Early childhood is a critical period of brain and cognitive development (Thomas & Johnson, 2008). Cross-sectional and longitudinal neuroimaging evidence in healthy humans has shown that from birth through adulthood, changes in whole-brain gray matter volume follow a non-linear trajectory with rapid increases in early life followed by gradual decreases during adolescence (Bethlehem et al., 2022; Brown et al., 2012; Lenroot et al., 2007; Mills et al., 2016; Narvacan et al., 2017; Remer et al., 2017; Rutherford et al., 2022). Increases in gray matter volume observed via magnetic resonance imaging (MRI) are thought to reflect biological processes such as neural proliferation, dendritic spine arborization, synaptogenesis, gliogenesis and maturation, maturation of the neuropil, and gyrification, while decreases are historically thought to reflect synaptic pruning and myelination, resulting in a change in the gray to white matter ratio (Deoni et al., 2015; Giedd & Rapoport, 2010; Raznahan et al., 2011; Remer et al., 2017). Gray matter maturation is spatiotemporally variable: parietal and inferior regions of the cortex reach peak volumes slightly earlier than frontal, temporal, superior, and subcortical regions (Bethlehem et al., 2022; Lenroot & Giedd, 2006; Lenroot et al., 2007; Tanaka et al., 2012; Uematsu et al., 2012; Zhang et al., 2021). However, only a few neuroimaging studies have been conducted spanning early childhood, which is a critical development period. Furthermore, there is a lack of evidence on brain development in early childhood from longitudinal studies, which are necessary for capturing true development trajectories (Di Biase et al., 2023). A more thorough understanding of typical gray matter development patterns in early childhood would help provide a basis for identifying atypical trajectories.

Sex differences in brain volumes have been widely reported in previous research (Cosgrove et al., 2007; De Bellis et al., 2001; DeCasien et al., 2022; Gennatas et al., 2017; Giedd et al., 2012; Goldstein et al., 2001; Gur et al., 2002; Kaczkurkin et al., 2019; Lenroot & Giedd, 2010; Lenroot et al., 2007; Lotze et al., 2019; Sanchis-Segura et al., 2019; Tanaka et al., 2012), which robustly show that males, on average, have larger gray matter, white matter, and total brain volumes than females across the lifespan. Previous research has also indicated that males have a more protracted development trajectory than females across late childhood and adolescence (De Bellis et al., 2001; Giedd et al., 2009; Lenroot et al., 2007; Tanaka et al., 2012). Notably, studies that examined gray matter volume as a proportion of total brain or intracranial volume suggest that females have proportionally more gray matter than males (Goldstein et al., 2001; Gur et al., 2002). Given the high degree of correlation between sex and total brain volume, both raw volume and normalized gray matter volume should be considered when testing for developmental sex differences (Giedd et al., 2012; Mills et al., 2016). While the presence of sex differences in brain structure during adolescence and adulthood is well established, it remains unclear whether those sex differences are already present in early childhood. Examining the timing with which brain sex differences emerge may help explain mechanisms underlying sex-differential prevalence rates for mental health and neurodevelopmental disorders (Copeland et al., 2014; Gogos et al., 2019; Halladay et al., 2015; Mathes et al., 2019; Maughan et al., 2013; Mazure & Swendsen, 2016; Merikangas & Almasy, 2020; Ramtekkar et al., 2010; Rutter et al., 2003; Smink et al., 2012).

The present study aimed to characterize longitudinal changes in regional gray matter volume across early childhood (ages 2–8 years) and compare development trajectories between males and females. We hypothesized that most group-level regional volume trajectories would show a quadratic increase in volume over time with a slowing of growth towards 8 years (inverted-U shaped trajectory). Additionally, we hypothesized that there would be significant effects of sex, both main effects (with males having larger total volumes) and interactions (with males having slower development than females). Regarding proportional volume, we hypothesized that global and regional gray matter volume would decrease over time and that sex effects would diminish when normalizing gray matter volume as a percent of total intracranial volume. While it is increasingly common to report additional gray matter measurements such as cortical thickness and surface area, which are more closely related to cytoarchitectural changes than gray matter volume alone, this manuscript focuses on gray matter volume for several reasons. First, we report development trajectories for both cortical and subcortical regions, the latter of which only volume can be reliably reported. Second, volume has been the most widely reported gray matter metric in previous studies, and thus reporting volume in this novel sample simplifies connecting our results to previous research.

This study was approved by The University of Calgary Conjoint Health Research Ethics Board (CHREB; REB13-0020). The present sample included 123 children from the prospective longitudinal Calgary Preschool MRI Dataset (Reynolds et al., 2020). The cohort consists of healthy children with no physical, neurological, or psychiatric diagnoses who were born at >35 weeks gestation. Children were recruited to the study between ages 2 and 5 years and completed MRI scans at 6–12 month intervals (N = 393 scans, ~3.2 scans/subject; Fig. 1) for a full age range of 1.97–8.04 years. 90% of parents were married/common law, 84% of families identified as White, 87% of mothers had an undergraduate degree or more, and 70% of family incomes were at or above the median income for the city of Calgary, Alberta, Canada (Table 1). MRI data were collected at the Alberta Children's Hospital on a research-dedicated 3 T GE MR750w MRI scanner with a 32-channel head coil. The MRI protocol included T1-weighted anatomical images (FSPGR BRAVO with 0.9 × 0.9 × 0.9 mm resolution, 210 axial slices, TR = 8.23 ms, TE = 3.76 ms, flip angle = 12 degrees, matrix size = 512 × 512, and inversion time = 540 ms). During T1 acquisition, children were awake viewing a movie of their choice or sleeping naturally.

Images were initially assessed for quality at the scanner during acquisition; sequences were repeated if necessary and if time permitted. Images were also examined for motion and those with major motion artifacts were excluded. During processing, N4 bias-corrected images (Cox, 1996; Cox & Hyde, 1997) were resampled to a voxel size of 1 mm³ in preparation for multi-atlas segmentation combined with cortical reconstruction using implicit surface evolution (MaCRUISE; Huo, Carass, et al., 2016; Huo, Plassard, et al., 2016; Huo et al., 2017). MaCRUISE integrates the processes of cortical reconstruction and multi-atlas segmentation to produce reliable and consistent cortical surface parcellations in anatomical agreement with brain segmentations (Huo, Carass, et al., 2016; Huo, Plassard, et al., 2016; Huo et al., 2017).

MaCRUISE has systematically been shown to reduce spatial inconsistencies between volumetric segmentation and surface parcellations in aging adult populations as compared to Freesurfer (Huo, Plassard, et al., 2016). We observed similarly incorrect surfaces in our pediatric data when processed with Freesurfer and subsequently decided to process all T1-weighted images used in this study with MaCRUISE. In the MaCRUISE pipeline (Huo, Plassard, et al., 2016), skull and dura stripped images are subject to both multi-atlas segmentation of 132 regions (Asman & Landman, 2012, 2013; Bermudez et al., 2020; Klein et al., 2010) and TOpology-preserving Anatomical Segmentation (TOADS) fuzzy membership segmentation (Bazin & Pham, 2008). MaCRUISE then fuses the rigid multi-atlas and TOADS segmentations, resulting in a full cerebrum segmentation comprising gray matter and white matter regions. To achieve a cortical reconstruction consistent with the segmentations, MaCRUISE applies multi-atlas anatomically consistent gray matter enhancement (MaACE; Han & Fischl, 2007; Sethian, 1999) to the gray matter component while applying a topology correction to the white matter component (Han et al., 2001, 2002). These refined gray and white matter segmentations form the outer and inner surfaces of the reconstructed cortex, respectively. Lastly, to resolve any remaining disagreements between the multi-atlas segmentation and reconstructed surfaces, MaCRUISE refines boundaries in the multi-atlas segmentation using the inner and outer cortical surfaces (Huo, Plassard, et al., 2016). While MaCRUISE is a surface-based volume pipeline (rather than voxel-wise), vertex-wise measurements for cortical thickness and surface area are not readily available as outputs from this pipeline and require further processing, outside the scope of this present study. We extracted the refined segmentations (in mm³) for analyses of absolute total gray matter volume and absolute regional volume. We additionally computed proportional volume by dividing each region’s volume by total intracranial volume (ICV) and multiplying by 100, resulting in the regional volume expressed as a percent of ICV.

Trained raters checked each individual segmentation image for accuracy, overlayed in each scan’s native T1-space, and assigned them a quality score of 1 (poor), 2 (unsatisfactory), 3 (satisfactory), and 4 (excellent). Segmentations with a quality score < 3 were manually edited and reintroduced to the MaCRUISE pipeline at the segmentation fusion step, in place of the original rigid multi-atlas segmentation. Edited segmentation outputs were reassessed for quality and included in the analysis if the resulting segmentation obtained a quality score of 3 or 4. Of the 452 successfully acquired T1-weighted images considered for this study, 393 (87%) had sufficient quality of both the T1-weighted image and resulting MaCRUISE segmentation. Of the excluded scans, 20 were the subject’s first session, 15 were the subject’s second session, 3 were the third, 4 were the fourth, 8 were the fifth, 4 were the sixth, and the remaining 4 scans were from the seventh timepoint or later. Only 1 subject was completely excluded from the study due to quality concerns. Spaghetti plots of raw volume measurements over time were visually inspected to ensure plausibility of intra-individual change in volume between scans.

To determine the best fitting polynomial trajectory of development, we successively fit and compared nested mixed-effects models of increasing complexity with subjects allowed a random intercept (R package lme4; Bates et al., 2015). We compared the following three models for each region:

where $Yij$ = j^th volume measurement for the i^th subject, $xij$ = the subject’s age at time of scan, $B1$ = coefficient for age, $B2$ = coefficient for age², $B0i$= subject-specific y-intercept, and $ϵij$= random error.

We determined the best fitting model for each brain region by selecting the one with the lowest Akaike Information Criterion (AIC; Sakamoto et al., 1986), and Bayesian Information Criterion (BIC; Neath & Cavanaugh, 2012), two loss metrics for model selection which also account for model complexity (Cox & Vladescu, 2023). While both AIC and BIC penalize a model for every additional parameter to reduce the risk of over-fitting, the penalty in BIC is greater. In this study, with a limited set of models compared at once, we deemed the risk of over-fitting to be low; In cases where AIC and BIC disagreed, we preferred the model selected by AIC. In this study, AIC and BIC selected the same model in ~75% of regions. Additionally, we calculated the Akaike weight for each model tested. The Akaike weight is a probability between 0 and 1 that a given model is the best at minimizing Kullbach-Leibler discrepancy among a set of models (Wagenmakers & Farrell, 2004). Akaike weights provide a means to compare the strength of evidence in favor of one model, as compared to another. We fit trajectory models for absolute total gray matter and absolute regional volumes (mm³) as well as for proportional total and regional gray matter volume expressed as a percent of ICV. For each best fitting model, we extracted the percent change in volume from baseline to peak, age at peak volume, and predicted volume at ages 2 and 8.5 years. We additionally determined partial eta squared (η_p²) for all age effects to examine the magnitude of effect that age/time has on developmental changes in gray matter volume. η_p² characterizes the proportion of variance in the dependent variable explained by an independent variable while accounting for other covariates in a model. Magnitude of effect size conveyed by η_p² is typically benchmarked as follows: small = 0.0099, medium = 0.0588, and large = 0.1379 (Cohen, 2013). Regions with non-null trajectories were retained for further analysis.

In a second step, using the best fitting non-null regional trajectories described above (i.e., final model was linear or quadratic), we created mixed-effects models to characterize the effects of child sex on gray matter development.

where $Yij$ = j^th volume measurement for the i^th subject, $xij$ = the subject’s age at time of scan, $Sij$ = subject sex (male = 0, female = 1), $B1$ = coefficient for age, $B2$ = coefficient for age², $B0i$= subject-specific y-intercept, $B3$ = coefficient for sex main effect, $B4$ = coefficient for age-by-sex interaction, $B5$ = coefficient for age²-by-sex interaction, and $ϵij$= random error.

To correct for multiple comparisons, we grouped p-values for like model components (i.e., all p-values for age-by-sex interaction term, all p-values for age²-by-sex interaction term, and all p-values for main effect of sex) and applied a false discovery rate (FDR; Benjamini & Hochberg, 1995) correction to each group of tests (absolute and proportional volume, main effect or interaction of sex) with a threshold of q < 0.05. When full models including age- or age²-by-sex interactions showed no significant effects of sex (q > 0.05), models were reduced (interactions removed, main effect of sex retained) and rerun. We examined residual distribution plots for all reported models to confirm adequacy of model fit.

Absolute total gray matter volume had an inverted U-shaped quadratic trajectory, increasing non-linearly by 8% between 2 and 6.85 years, or between 1% and 4% annually (Tables 2, S1; Fig. 2). Proportional gray matter volume was 57.7% of ICV at age 2 and 55.8% of ICV at age 8 years. The trajectory followed a slightly inverted U-shape with a peak at 3.67 years (Table 2, S1; Fig. 2). Males had significantly larger absolute total gray matter volume than females (Table S2; Fig. 2), although this effect disappeared when normalizing by ICV (Table S3, Fig. 2). This pattern was confirmed by evaluation of the AIC values and weights; the best fitting trajectory model for total gray matter volume included a main effect of sex and evidence strongly favored this model (AIC = 9281.99, w = 0.84) and proportional volume was best fit by a trajectory including no effects of sex (AIC = 1435.26, w = 0.64). We found no significant sex differences in trajectory shape for either absolute or proportional total gray matter volume (Table S2, S3, Fig. 2).