Typical development of the human fetal subplate: Regional heterogeneity, growth, and asymmetry assessed by in vivo T2-weighted MRI

Significance

The subplate (SP) is a transient fetal brain compartment that provides an early foundation for downstream cerebral development. Using fetal MRI, we show that SP morphology is regionally heterogeneous and that its growth follows distinct regional trajectories. Principal spatiotemporal patterns differentiate early- from later-developing regions and partially recapitulate sensory-to-association hierarchies of cortical maturation. SP also exhibits dynamic hemispheric asymmetries as early as the second trimester, with some amplifying as potential precursors of postnatal lateralization and others reversing with age, emphasizing the importance of evaluating developmental trajectories rather than single time points. This framework positions SP morphology as an accessible early marker of fetal brain organization with potential relevance for later neurodevelopment.

Abstract

The subplate (SP) is a transient fetal brain compartment supporting neuronal migration, axonal ingrowth, and early cortical activity, yet the dynamics of its regional development remain poorly understood in vivo. Using T2-weighted fetal MRI of 68 typically developing fetuses (22 to 32 wk gestational age, GA), we developed a semiautomated pipeline to quantify regional SP morphology (thickness, surface area, and volume). SP characteristics scaled strongly with GA and residual brain volume and showed marked regional differences. After correcting for geometric confounds, regional variation of SP thickness persisted, with highest values in parietal and perisylvian regions, suggesting that SP thickness may serve as a sensitive marker of intrinsic developmental differences. Between the late 2nd and early 3rd trimesters, mean SP thickness increased by 39.2% with large variation across regions (±11.0 SD), whereas surface area growth was more uniform (64.3% ±0.7 SD). Continuous growth trajectories clustered into distinct spatiotemporal profiles: early-developing regions (e.g., pericentral and medial occipital cortices) contrasted with later-developing regions (prefrontal, temporal, and parietal cortices). These patterns partially recapitulate primary-to-association, medial-to-lateral, and posterior-to-anterior maturational hierarchies, pointing to organized developmental programs. SP development also showed region-specific hemispheric asymmetries, including leftward thickness and volume asymmetry in the superior temporal and precentral gyri. Some asymmetries amplified, others attenuated or reversed with age, suggesting both transient states and potential precursors of postnatal lateralization. Together, these findings provide a framework for regional SP quantification and position SP morphology, particularly thickness, as a promising early biomarker that might link fetal SP changes to subsequent cortical development and neurodevelopmental outcomes.

Fetal brain development unfolds as a cascade of rapid and interrelated events that lay the foundation for mature brain organization and function (1, 2). The importance of this period for optimal neurodevelopment is increasingly recognized, based on accumulating evidence of lifelong functional consequences associated with adverse in utero environments or early-life events such as prematurity (2, 3) and in utero origins suggested for various neurodevelopmental disorders (4, 5).

A characteristic feature of the fetal brain from mid-gestation onward is the presence of transient compartments, most notably the subplate (SP) (6). Foundational insights into SP development and function derive from animal models (7–11) and postmortem human studies using histology and ex vivo MRI (6, 12–15). These investigations reveal SP’s complex architecture, including its sublaminar organization and the morphological and molecularly diverse populations of SP neurons, radial and tangential migratory streams, glial precursors, signaling molecules, and an abundant extracellular matrix (ECM) (16–18). Crucially, the SP serves as a transient convergence zone for multiple afferent systems, including thalamocortical, basal forebrain, and early corticocortical projections (19–22). The initial synaptic connectivity established within the SP contributes to early functional network activity (23, 24). At its peak, the SP can be up to four times thicker than the overlying cortical plate (CP), reflecting its high developmental load. As afferent progressively relocate to the CP, the SP gradually dissolves during the last third of gestation, marking a key transition toward more mature cerebral organization (25). Crucially, the SP is not merely a transient compartment but actively participates in key corticogenic processes, shaping the formation of permanent neural circuits and cortical columnar organization by influencing CP differentiation and synaptogenesis. Through these roles, the SP likely contributes to cortical arealization and has been implicated in the mechanics of cortical folding (26) [for review (16)]. Thus, detailed in vivo characterization of SP morphology might offer a unique window into the early spatial patterning of the developing human brain, with important implications for understanding cortical developmental trajectories.

Importantly, atypical features of the SP have been hypothesized in several neurodevelopmental disorders, including cerebral palsy, autism spectrum disorder (ASD), attention deficit hyperactivity disorder (ADHD), and schizophrenia (27, 28). These underscore the potential of SP characteristics to serve as early biomarkers of later neurodevelopmental dysfunction. However, accurate interpretation of such deviations requires a detailed understanding of typical SP development, which itself is highly dynamic and marked by substantial spatial and temporal heterogeneity in its emergence, expansion, and dissolution (11). This pronounced morphological variability is not only biologically significant but also presents unique computational challenges for SP quantification, which may partly explain why systematic in vivo assessment of the SP has been limited.

Increasingly, recent advances in in utero MRI, driven particularly by advances in rapid, motion-tolerant acquisition methods [e.g., HASTE (29), echo-planar FLAIR (30)] and dedicated fetal brain processing pipelines (31–34), are beginning to address these challenges, enabling in vivo study of the SP (35, 36). These technological advances have the potential to support anatomically precise cross-sectional as well as longitudinal investigations of fetal brain tissue compartments. The increasing availability of larger fetal datasets permits capturing developmental changes with higher temporal resolution. For example, several studies have quantified SP development using global metrics such as total volume (37, 38), T2 signal intensity (30), and microstructural properties (39, 40). On a regional level, expansion of SP thickness has been described particularly in sensorimotor and occipital regions early in gestation (41). Subsequent studies have extended these findings into later gestation, mapping regional SP volumes and linking them with sulcal emergence and gene expression gradients (42–44).

However, gaps remain in this literature. Most prior in vivo studies have either focused on global SP metrics or included the subjacent intermediate zone (IZ), which limits regional and anatomical specificity. Although some studies have noted regional heterogeneity [in SP thickness (41) or volumes (42–44)], few have assessed SP development using multiple morphological dimensions (e.g., thickness, surface area, volume), although each dimension may exhibit distinct developmental signatures while reflecting interrelated biological processes. Moreover, SP assessment has been constrained by limited subject availability and also the time-consuming nature of manual segmentation or manual correction of registration-based SP labels. To address these limitations, we applied an in-house automated fetal brain processing pipeline, which enables robust segmentation of the SP and extraction of regional surface-based and volumetric features across the cerebrum (33, 34, 45). This approach allows detailed in vivo characterization of SP development.

We focused on the 22 to 32 wk gestational age (wGA) period. During this window, the SP appears as a distinct hyperintense band between the CP and IZ on T2-weighted MRI, primarily due to its ECM content (18, 35). Although the SP is histologically evident from 15 to 17 wGA and functionally significant by approximately 20 wGA (46, 47), earlier imaging is limited by small brain size and motion artifacts. By the upper limit of this window (~32 wGA), contrast between the SP and IZ progressively decreases due to ECM dissolution and afferent relocation to the CP, rendering the SP indistinguishable from the IZ (30, 48). Thus, the 22 to 32 wGA period offers optimal contrast for SP segmentation. Incidentally, this window also overlaps with peak SP prominence (before its progressive dissolution) (49), coincides with intense proliferative and migratory events and axonal growth (49), and aligns with the lower limit of viability in preterm birth (50), making it both biologically and clinically relevant.

In this study, we characterized regional heterogeneity in SP thickness between 22 to 32 wGA to test whether inherent differences exist across the cerebrum. We then assessed regional SP development during this period by evaluating relative changes in regional SP thickness, surface area, and volume between the late 2nd (22 to 27 wGA) and early 3rd trimester (>27 to 32 wGA) periods, to identify regions of peak developmental activity. We also clustered regional developmental trajectories to identify common growth profiles. Finally, given the important hemispheric lateralization of adult brain organization (51–53) and evidence for prenatal origins of this lateralization (40, 43, 54–58) we examined interhemispheric asymmetries across regional SP characteristics to explore whether lateralized features were already apparent during this period and how they evolved with gestational age (Fig. 1A).

Fig. 1.

(A) Schematic overview of the MRI processing pipeline. (B) Visualization of the final set of regions used in the study, along with the corresponding color scheme. (C) Summary of key analyses.

Results

Gestational Age, Brain Volume, and Regional Variation Modulate SP Characteristics.

Using a near-automated processing pipeline, we quantified SP thickness, surface area, and volume, enabling detailed assessment of developmental dynamics. First, we examined the global effects of GA, residual brain volume (supratentorial brain volume adjusted for GA), and sex on SP morphology. ANCOVA modeling revealed strong positive association with GA for whole-brain SP thickness (F = 393.553, η² = 0.774), surface area (F = 6,167.471, η² = 0.950), and volume (F = 2,823.253, η² = 0.927) (all P < 0.001; SI Appendix, Figs. S1 and S2 and Table S1). Residual brain volume was independently associated with SP thickness (F = 52.507, η² = 0.103), surface area (F = 248.569, η² = 0.038), and volume (F = 248.569, η² = 0.052) (all P < 0.001), confirming scaling of SP morphology with overall brain size. Sex effects were minimal, likely because residual brain volume accounted for most of the sex-related variance (residual brain volume: male vs. female, t = 4.95, df = 63.11, P < 0.001, Cohen’s D = 1.21), with only SP volume showing a significant but negligible sex effect (F = 5.594, P = 0.021, η² = 0.002), and interactions with GA were also negligible (F = 13.120, P = 0.001, η² = 0.002). Hence, sex was not considered in further analyses.

We next examined thickness variation across SP regions and hemispheres using repeated-measures ANCOVA. Cortical folding depth (“SP depth”) is geometrically related to SP thickness, confirmed by a significant within-subject effect (F = 190.129, P < 0.001, η² = 0.016) and a modest interaction with region (F = 2.321, P = 0.002, η² = 0.003) (full statistical results are provided in SI Appendix, Table S2). Substantial regional differences in SP thickness persisted after correcting SP thickness for regional SP depth to account for these geometric influences of outer SP undulations on thickness (F = 111.956, P < 0.001, η² = 0.271; Fig. 2A and SI Appendix, Fig. S3 and Table S3). Small but significant interactions were also observed for region × GA (F = 9.052, P < 0.001, η² = 0.022), region × hemisphere (F = 26.278, P < 0.001, η² = 0.025), and their three-way interaction (F = 4.215, P < 0.001, η² = 0.004), suggesting that growth rates of SP thickness vary regionally and show hemispheric asymmetry. These results indicate that SP thickness heterogeneity cannot be explained solely by global brain growth or differences in cortical folding and may reflect regional differences in axonal ingrowth, cellular composition, or extracellular matrix volume.

Fig. 2.

Regional variability of SP thickness and relative change of SP characteristics between the late 2nd and early 3rd trimesters. (A) Regional variability in SP thickness after accounting for covariates: sulcal depth within subjects, and GA and residual brain volume within ROIs. (B) Mean relative change in SP thickness, surface area, and volume projected onto SP surface maps, illustrating spatial patterns of growth. (C) Regional distribution of relative growth, with dots indicating mean values and error bars representing CI. Regions outlined by boxes show significant left–right differences in relative growth across the period (FDR-corrected across regions within each metric) suggesting developmental asynchrony.

Similar modeling of SP surface area and volume reveal small but significant interactions for region × GA (surface area: F = 5.550, P < 0.001; volume: F = 23.520, P < 0.001) and region × hemisphere (surface area: F = 26.780, P < 0.001; volume: F = 29.731, P < 0.001), along with their three-way interactions (surface area: F = 3.180, P < 0.001; volume: F = 6.808, P < 0.001), though effect sizes were small (η² < 0.009). These interactions indicate subtle region- and hemisphere-specific growth patterns, but interpretation of absolute biological differences is limited, as the size of parcellated regions depends on the predetermined parcellation scheme. Overall, these analyses demonstrate that SP growth is closely coupled to overall brain development and size, with considerable regional and hemispheric variability, and suggest that thickness may provide a sensitive measure of region-specific heterogeneity, motivating subsequent analyses of regional growth patterns and hemispheric asymmetries.

Analysis of Regional SP Growth Identifies Distinct Spatiotemporal Developmental Patterns across 22 to 32 wGA.

We next characterized developmental dynamics of SP growth as a relative regional growth between the late 2nd (22 to 27 wGA) and early 3rd (27 to 32 wGA) trimesters, after adjusting regional values for the same confounders as before within groups. SP thickness, surface area, and volume increased across all regions, but the extent of this growth varied by metric and region Fig. 2 B and C and SI Appendix, Table S4). SP volume showed a mean growth of 81.3% and a large SD of 16.6. The inferior temporal, fusiform, and middle frontal gyri showed the largest volume increases (up to +98%). In contrast, surface area growth was more uniform with a low SD (mean growth 64.3% ±0.7), peaking in the lingual gyrus (+65%). Thickness changes were also heterogeneous (mean growth 39.2% ±11.0 SD), with the greatest increases in the paracentral lobule, right superior parietal cortex, and left precentral and lingual gyri (+49%). These patterns suggest that SP growth is regionally orchestrated, with certain regions undergoing more rapid expansion during mid-gestation, potentially supporting early cortical circuit establishment.

To elucidate dominant spatiotemporal patterns, Ward’s hierarchical clustering was applied to continuous regional growth profiles. Four clusters were identified for SP thickness (Silhouette = 0.513), and three clusters each for surface area (Silhouette = 0.406) and volume (Silhouette = 0.361) based on Silhouette scores and dendrogram inspection (Fig. 3 and SI Appendix, Fig. S4 and Table S5). Cluster-averaged profiles revealed differences in peak growth rate, timing, and cumulative growth (Table 1).

Fig. 3.

Clustering of SP developmental trajectories between 22 to 32 wGA. (Top row) Results of hierarchical clustering (Ward’s method) applied to the regional developmental profiles for SP thickness, surface area, and volume. (Bottom row) Cluster-averaged growth profiles showing mean local relative growth rates for each cluster over time and associated CI. Vertical lines indicate the time point of peak growth for each cluster.

Table 1.

Summary of developmental trajectories for subplate (SP) clusters

	Peak rate (%/week)	Peak time (week)	AUC (%)	Growth density (%/week)
SP thickness
Cluster T1	249.48	22.05	119.75	4.83
Cluster T2	160.61	22.05	87.03	3.03
Cluster T3	110.01	31.35	73.44	3.73
Cluster T4	88.78	22.05	50.39	1.74
SP surface area
Cluster SA1	139.35	30.25	91.99	5.37
Cluster SA2	133.30	29.75	98.07	4.16
Cluster SA3	129.13	29.65	98.88	3.61
SP volume
Cluster V1	268.54	22.05	116.38	5.52
Cluster V2	139.17	27.75	121.09	2.19
Cluster V3	150.06	29.35	107.92	5.15

For each cluster, the table reports peak relative growth rate (%/week), peak timing (wGA), cumulative growth (AUC, computed as the area under the relative growth rate curve), and growth density (SD of the smoothed growth curve).

For instance, SP thickness clusters T1 and T2 peaked early: T1 (e.g., right superior frontal, bilateral paracentral lobule, lingual gyri) showed the highest peak rate (249.5%/week) and cumulative growth (AUC = 119.8%), while T2 (areas around central sulcus, left temporal lobe) had more moderate growth density (i.e., temporal concentration of growth; 3.03%/week) and cumulative growth (AUC = 87.0%). T3 (mainly parietal and parts of the temporal cortex) peaked the latest (~31.35 wGA) with slower growth (110%/week). T4 (lateral occipital, and parts of right temporal lobe, precuneus) peaked early but showed the lowest growth density (1.74%/week) and cumulative growth (AUC = 50.4%). Surface area clusters peaked more uniformly around 30 wGA. The mainly frontal cluster SA1 showed highest growth density (5.37%/week) and latest peak (30.25 wGA), while SA2 and SA3 showed similar cumulative growth (AUC ~ 98%) with lower growth density (<4.16%/week). Volume clusters displayed more heterogeneity: V1 peaked early with the highest growth rate (268.5%/week) and growth density (5.52%/week); V2 (mainly parietal) peaked later (27.75w) with the highest cumulative growth (area under the relative growth curve [AUC] = 121.1%) and moderate growth density (2.19%/week); V3 (mainly frontal) peaked last (29.35w) with high growth density (5.15%/week) but lower cumulative growth (AUC = 107.9%).

SP Characteristics Show Robust, Region-Specific Hemispheric Asymmetries That Evolve between the Late 2nd (22 to 27 wGA) and Early 3rd (27 to 32 wGA) Trimesters.

Comparisons of homotopic regions for SP thickness, surface area, and volume with paired t-tests revealed significant left–right differences after adjusting for confounders (GA, residual brain volume, and additionally regional SP depth in case of SP thickness), with asymmetry indices (AIs) confirming consistent lateralization patterns (Fig. 4A). Rightward asymmetries were observed in middle frontal, inferior frontal, and postcentral gyri across all SP characteristics (T = 3.794 to 13.519, Cohen’s D = 0.28 to 0.61, P < 0.001–0.01, FDR-corrected across regions within metric). Additional rightward asymmetries were seen for example in lingual gyrus and superior parietal cortex (thickness, volume), fusiform gyrus (thickness), and paracentral lobule (volume). The precuneus consistently showed leftward asymmetry across all metrics (T = –14.386 to –4.391, Cohen’s D = 0.42 to 1.59, P < 0.001). Leftward asymmetries were also present in cuneus and inferior parietal cortex (volume and thickness), precentral gyrus (volume), and supramarginal gyrus (surface area). Some regions, including the lateral occipital cortex, superior frontal, and middle temporal gyri, exhibited metric-specific asymmetries, highlighting that thickness, surface area, and volume each capture distinct aspects of SP development.

Fig. 4.

Asymmetry analysis. (A) Asymmetry evaluations of regional SP characteristics: thickness, surface area, and volume. The Upper row displays mean asymmetry index values projected onto cortical surface maps. The Lower row presents the distribution of asymmetry indices across regions. Only regions exhibiting statistically significant asymmetries (P < 0.05, FDR-corrected across regions within metric) are shown in color on surface plots. Regions with significant rightward asymmetry are color-coded in red, while those with significant leftward asymmetry are shown in blue. (B) Developmental changes in hemispheric asymmetry between the late 2nd and early 3rd trimesters. For each region, the asymmetry index (AI) is shown at late 2nd (L2) and early 3rd (E3) trimesters. Colored dots indicate significant lateralization (t test vs. 0, FDR-corrected across regions within metric): red for rightward, blue for leftward, gray for nonsignificant. Dots represent mean AI; vertical lines show 95% CI. Lines connecting L2 and E3 indicate change in AI between the periods: red for rightward, blue for leftward shifts (Z-test on difference between timepoints, FDR-corrected across regions within metric), gray for nonsignificant changes.

Analysis of changes in AIs over time revealed that asymmetries can emerge, amplify, attenuate, or even reverse in a region-specific manner between the late 2nd and early 3rd trimesters (Fig. 4B and SI Appendix, Table S6). For example, the lingual gyrus showed a significant increase in rightward SP volume asymmetry (ΔAI = 5.08, Z = 2.35, P = 0.036), while the lateral occipital cortex, paracentral lobule, and postcentral gyrus developed new rightward asymmetry by mid-gestation. The precuneus maintained strong leftward asymmetry with a significant volume AI increase (ΔAI = 5.12, Z = 2.46, P = 0.035). Dynamic changes included attenuation of leftward volume asymmetry in the inferior parietal cortex due to rightward-biased growth (ΔAI = 6.07, Z = 2.40, P = 0.035), reversal of the superior frontal gyrus from leftward to rightward asymmetry (late 2nd: AI=–4,18, P = 0.043; early 3rd: AI = 3.13, P = 0.003), and reversal of thickness asymmetry in the inferior temporal gyrus from rightward to leftward (late 2nd: AI=10.27, P = 0.006; early 3rd: AI = −4.19, P = 0.048).

Cluster spatial analysis indicated moderate bilateral symmetry across SP characteristics, strongest in volume clusters and weakest in thickness clusters, consistent with thickness’s greater variability in its spatial and temporal developmental profiles (Fig. 5).

Fig. 5.

Matrices comparing bilateral asymmetry across clusters. Each cell shows the normalized symmetry between left and right hemisphere regions within or across clusters (higher values indicate greater symmetry). Diagonal entries thus reflect the degree of symmetry for homotopic regions assigned to the same cluster.

Discussion

We present a framework for in vivo quantification of regional SP morphology using T2-weighted fetal MRI. Taking advantage of 3 T acquisitions, which can achieve higher signal-to-noise ratio and spatial resolution than more commonly used 1.5 T systems (59), this approach enables reliable characterization of SP thickness, surface area, and volume between 22 to 32 wGA. Our findings demonstrate that SP morphology is both regionally heterogeneous and develops along region-specific trajectories. Relative growth analyses and hierarchical clustering reveal principal spatiotemporal patterns, differentiating early- from later-developing regions. SP also exhibits dynamic, region-specific hemispheric asymmetries emerging as early as the second trimester. Together, these findings highlight the complex interplay of intrinsic regional properties, growth timing, and lateralization, underscoring SP’s role as an early scaffold for downstream cortical development.

Regional Heterogeneity in SP Thickness.

SP thickness showed pronounced regional differences that persisted after controlling for GA, residual brain size, and SP depth. Parietal and perisylvian regions exhibited the greatest thickness, consistent with histological and imaging studies (22, 35, 41, 60). These differences likely reflect multiple factors, including differential afferent input, for example, early thalamocortical and dense associative connectivity linked to higher SP thickness in perisylvian and parietal areas (11, 46, 61) or higher somatosensory thalamic (22) compared to motor afferents that could explain greater SP thickness in the postcentral vs. precentral gyrus. Intrinsic regional factors, including previously described variations in neuron density, molecular phenotype, glial composition, and ECM, also likely further modulate thickness directly or indirectly through further modulation of afferent targeting and synaptic remodeling (16, 17, 25, 62) and merit further study to elucidate the drivers of inherent differences in SP thickness across regions. Studies using diffusion imaging, particularly in conditions altering afferent input (e.g., thalamic injury, callosal agenesis) will be useful in this context.

Importantly, our findings complement prior work showing spatial patterning of other transient fetal compartments (1). Combined with previous observations of consistent regional CP/SP volume proportions across gestation, and links between CP and SP thickness and regional gene expression (42, 44), the evidence is consistent with the protomap hypothesis, suggesting that fundamental organization and regional identity may be determined early in development by intrinsic factors (63, 64). Interestingly, SP is not a passive relay but an active scaffold for early circuit formation. For example, SP neurons serve as pioneers scaffolding axonal pathways (65), with specialized SP corridors guiding thalamic axons described in occipital regions (66), and SP further influences cortical plate differentiation and synaptogenesis (23, 67, 68). Although we project cortical regions onto the underlying SP assuming a one-to-one correspondence, this relationship is likely more dynamic, shaped by tangential migration and regional developmental gradients (69) rather than a fixed columnar organization, in vivo SP thickness may serve as a relatively accessible quantifiable marker of early arealization with potential to predict later cortical development that deserves further attention.

Asynchronous Regional Growth between the Late 2nd and Early 3rd Trimesters.

Our results demonstrate asynchronous regional SP development, echoing prior histological and imaging work (70). Between the late 2nd and early 3rd trimesters, relative growth in SP thickness varied widely: Paracentral lobule showed the greatest increases, potentially reflecting the changes in its motor-related part aligning with late second trimester motor area maturation (70). Precentral gyrus thickening aligned with this in comparison to the postcentral gyrus which showed more modest change, possibly due to earlier somatosensory afferent arrival. Medial occipital regions (cuneus, lingual gyrus) thickened more than the lateral occipital cortex, consistent with earlier medial visual system maturation (44). These in vivo observations are consistent with postmortem MRI evaluations, which report relatively higher SP thickness in dorsolateral compared to dorsomedial regions as well as higher SP growth in lateral vs. medial areas (71). Association cortices such as inferior parietal and temporal areas, linked to human-specific evolutionary expansion (72, 73), also exhibited pronounced thickening, perhaps reflecting prolonged SP development in evolutionarily significant areas.

In contrast, surface area growth was more uniform, suggesting broadly isometric surface expansion during this gestational window. We speculate that global tangential expansion might be spatially constrained by adjacent regions, favoring greater radial thickening in areas with pronounced SP growth, particularly in the association cortex, to mitigate developmental demands and spatial competition. Such localized thickening may in turn alter regional biomechanics, potentially influencing downstream folding patterns, especially relevant to secondary and tertiary sulci that emerge after 32 wGA (74) and to the expanded surface area of the association cortex (72). Prior work implicates subcortical stiffness and growth gradients in folding mechanics (75, 76), and imaging evidence suggests that SP changes precede sulcal formation (39, 77). Whether regional SP thickening modulates local stress and contributes to species-specific folding remains unresolved and will require integration of high-resolution fetal diffusion MRI, biomechanical modeling, and comparative developmental studies. Moreover, we note that regional growth between the late 2nd and early 3rd trimesters involved relatively small numbers (N = 29 to 39), which may limit robustness of confounder-corrected estimates. Rigorous QC and agreement with prior work support the reliability of observed patterns, though larger longitudinal cohorts are needed to confirm them.

Dominant Growth Profiles from Clustering.

Hierarchical clustering of continuous regional trajectories revealed dominant developmental patterns for thickness, surface area, and volume. Early thickening clusters (T1: paracentral lobule, lingual gyrus; T2: pericentral, cuneus, left temporal regions) partially align with early sensory systems (13, 78), while later clusters (T3: parietal and lateral temporal gyri) followed more delayed thickening potentially typical of the association cortex (16). These patterns partially recapitulate sensory-to-association hierarchies observed in white matter connectivity or cortical microstructure (79) and may anticipate postnatal sensory-association gradients of functional organization (80–82). SP volume clusters showed similar “gradient” (early-growing V1: postcentral, middle temporal; followed by later, more steadily growing V2: parietal, lateral occipital, inferior temporal; and V3: frontal, medial occipital, superior temporal). Exceptions, such as the left middle temporal gyrus grouping with early-thickening regions, may reflect accelerated development of its posterior, more sensory-related subregions (83), underscoring the limitations of our relatively coarse parcellation scheme which might mask finer-scale heterogeneity. Differences between medial vs. lateral occipital regions (T1/T2 vs. T3), and paracentral vs. pre/postcentral gyri (T1 vs. T2) further suggest spatially organized development, suggestive of lateral-to-medial (and potentially posterior-to-anterior) sequences akin to those in other developing tissues (83). Moreover, previous work investigating regional signal intensity differences across the cerebral wall further supports a medial-to-lateral developmental gradient and suggests anterior-to-posterior variation evolving with gestational age, consistent with the spatial patterns identified in our clustering analyses (71). Overall, our results support spatially organized developmental programs within SP that reflect known developmental patterns and could anticipate future functional organization while interacting with asynchronous, region-specific processes, highlighting SP development as a critical entry point for understanding cortical development.

Asymmetries of SP Morphology and Their Developmental Dynamics.

Hemispheric asymmetry is a known feature of mature brain organization (52, 53), with some prenatal origins suggested by prior imaging (43, 54, 56). However, description of such asymmetries in transient fetal structures like the SP remains limited. Here, we report significant region-specific lateralization across SP characteristics. For example, the superior temporal gyrus showed leftward thickness and volume asymmetry, aligning with prenatal left temporal enlargement (54), while rightward surface asymmetry in superior and middle temporal gyri paralleled earlier right sulcal development (56). Precentral and postcentral gyri displayed opposing volume asymmetries: leftward in precentral [mirroring adult motor asymmetry (52)] and rightward in postcentral, the latter previously unreported observation. Frontal regions were mostly right-lateralized except the inferior frontal gyrus, partly diverging from previous fetal studies (43). Such variation likely reflects the individual heterogeneity (84), dynamic nature of SP development, and methodological limitations. For example, although we used a symmetric parcellation scheme, residual asymmetries may partly reflect parcel misalignment due to folding asymmetries, which could influence some region-specific estimates.

Additionally, asymmetry profiles evolved over gestation. Some regions (e.g., lingual gyrus, precuneus, pre/postcentral gyri) maintained or strengthened asymmetry, suggesting early subtle biases, including those in regional gene expression and molecular signatures, amplified over time (84). In contrast, other asymmetries attenuated or reversed during the evaluated period (e.g., inferior parietal cortex showing leftward volume but decreasing over time; attenuation of surface asymmetry in inferior temporal gyrus and paracentral lobule; or rightward-to-leftward reversal of thickness asymmetry in inferior temporal gyrus and the reverse for volume asymmetry in the superior frontal gyrus). These findings indicate that while some fetal SP asymmetries resemble patterns observed later in life and could represent early precursors of postnatal lateralization, their dynamic nature highlights the complexity of fetal SP lateralization. This cautions against direct extrapolation to adult patterns, particularly given extensive sensory-dependent perinatal cerebral remodeling, and underscores the importance of both timing assessments and evaluating developmental trajectories over extended periods.

Based on developmental profiles, homotopic regions largely clustered together for volume and surface area growth, and to a lesser degree for thickness, indicating symmetrical timing and growth dynamics despite morphometric asymmetry. We propose that this suggests that SP asymmetries arise from localized regional differences (85) superimposed on broadly symmetrical developmental programs. Given SP’s role in afferent integration and circuit scaffolding, early lateralization may bias later functional specialization, and its disruption could contribute to neurodevelopmental disorders in which altered asymmetries have been reported (e.g., autism, ADHD, dyslexia) (52, 86, 87). Determining which fetal asymmetries persist and which resolve will be essential for understanding the origins of these altered patterns, making longitudinal, multimodal studies of SP morphology particularly valuable.

Conclusion

Our in vivo characterization of SP between 22 to 32 wGA reveals spatially heterogeneous, temporally dynamic, and asymmetrically patterned development. Regional trajectories of thickness, surface area, and volume, along with clustering analyses, identify spatially organized programs that partially recapitulate known developmental hierarchies. Early, evolving asymmetries further demonstrate fetal lateralization. These results establish a robust framework for in vivo SP quantification, offering an early entry point to study how fetal SP development influences downstream cortical maturation, connectivity, folding, and neurodevelopmental outcomes, providing a foundation for future biomarker studies.

Materials and Methods

Subjects.

This study was conducted under a protocol approved by the Boston Children’s Hospital Institutional Review Board (IRB-P00008836), and written informed consent was obtained from all parents. Fetuses with congenital brain malformations, incidental MRI findings, or serious maternal medical conditions were excluded. The final cohort included MRI of 68 typically developing singleton fetuses (36 males, 32 females), scanned at mean 27.4 wGA, range: 22.0 to 31.4 wGA (SI Appendix, Fig. S5). Only scans before 32 wGA were included because SP contrast becomes indistinguishable from the intermediate zone on T2-weighted MRI beyond this age (30), as confirmed in our data by SP/IZ contrast approaching 0 toward 32 wGA (SI Appendix, Fig. S6). Future multimodal imaging incorporating complementary T1- and T2-weighted data (38), unavailable in our dataset, may extend these observations into late gestation. Mean maternal age was 32.6 y, range: 21.8 to 40.0. For developmental analyses, fetuses were grouped into late 2nd (22 to 27 wGA, N = 29) and early 3rd (>27 to 32 wGA, N = 39) trimester groups. A summary of cohort characteristics is provided in SI Appendix, Table S7.

Data Acquisition and Reconstruction.

MRI was performed on 3 T scanners: 42 scans on a Siemens MAGNETOM Skyra and 26 on a Siemens MAGNETOM Prisma (Siemens Healthineers, Germany). Imaging included repeated multiplanar T2-weighted half-Fourier single-shot turbo spin echo (HASTE) sequences optimized for fetal imaging. For each subject, HASTE stacks were acquired in axial, coronal, and sagittal orientations at least three times, totaling ~30 min of acquisition time. Parameters included TR = 1,400 to 1,600 ms, TE = 99 to 132 ms, 1 mm in-plane resolution, 2 to 3.5 mm slice thickness, and variable field of view adapted to fetal and maternal size (36). Acquisition at 3T provides improved signal-to-noise characteristics and achieves higher resolutions compared to more standard 1.5 T (59), supporting more reliable delineation of the subplate as a distinct transient compartment on T2-weighted imaging across gestation. Increased susceptibility to B1 field inhomogeneity and motion-related artifacts were addressed using a dedicated processing pipeline that included brain extraction (45), N4 bias field correction (88), and slice-to-volume reconstruction with NeSVoR (33), resulting in superresolved motion-corrected 3D T2w volumes at 0.5 mm isotropic resolution. Additional details on the HASTE stacks included in the reconstructions and the resulting reconstruction quality are provided in SI Appendix, Fig. S7.

SP Segmentation and Surface Extraction.

T2w volumes were registered to a 31w fetal brain template using rigid-body registration with scaling (7 dof) with FSL’s FLIRT (89, 90), as this template space is expected by our segmentation models. SP segmentation was performed with an ensemble of 2D attention-gated U-Net models trained on sagittal, coronal, and axial views, with aggregation and test-time augmentation, where the final label was selected based on aggregated probabilities across model predictions to improve robustness. This approach adapted a previously validated CP segmentation method (45) fine-tuned to include SP labels (91). Segmentations were visually inspected and manually corrected by a single trained rater to address any segmentation errors based on underlying T2-weighted images. Intrareader SP segmentation reliability was assessed in 15 randomly sampled subjects (mean 27.2 wGA, range 22.7 to 31.3) and demonstrated very high agreement (mean Dice coefficient ≈ 1, SD ≈ 0; mean volume difference 0.01%, SD 0.011). All segmentations in this study were corrected by this reader; additionally, a comparable interreader assessment against a second trained reader was performed also achieving high agreement (mean Dice coefficient 0.99, SD ≈ 0; mean volume difference 0.09%, SD 0.09) (SI Appendix, Fig. S8).

Outer SP surfaces (CP/SP boundary) were extracted using the CIVET marching cubes algorithm (92) using the ChRIS pl-fetal-surface-extract plugin (93) and postprocessed with Taubin smoothing (94) to suppress overfitting to the coarse voxel boundaries. Inner surfaces (SP/IZ boundary) were generated by inward deformation along subject-specific radial distance maps optimized according to the subject’s estimated gyrification index with empirically determined adaptive scheduling to accommodate diverse developmental morphologies (95). This approach thus preserves mesh topology and vertex correspondence between outer and inner SP surfaces. Surfaces were resampled to standardized meshes of 81,920 triangles and 40,962 vertices. Surface quality was evaluated visually and quantitatively via smoothness error (mean curvature difference between each vertex and its neighbors) and boundary distance error (Euclidean distance of vertex to nearest boundary voxel) (SI Appendix, Fig. S9).

Regional Parcellation.

21 bilateral regions were delineated by adapting original 34-region Desikan-Killiany atlas (96) to the fetal brain as proposed in ref. 44 with three exceptions: 1) separate superior, middle, and inferior temporal gyri; 2) subdivision of pericalcarine into cuneus and lingual gyrus; and 3) merging gyrus rectus with orbital frontal gyrus (Fig. 1B and SI Appendix, Table S8). Parcellations were manually defined on a 29w surface template (90, 97) and projected to individual SP surfaces using 2D surface-to-surface spherical warping (98, 99). Resulting parcellations were visually inspected to verify gross anatomical alignment and to identify obvious registration failures. This approach provided consistent regional delineations (SI Appendix, Fig. S10) and has been successfully applied in prior studies from our group (97, 100, 101). Surface labels were then mapped to volumetric space using a ribbon-constrained projection adapted from Connectome Workbench (v2.0.1), restricted to the SP label to yield volumetric SP regions. Regions with poor data quality were excluded: orbitofrontal cortex (due to frequent fronto-ventral blur on T2w), and pericingular pole regions: cingulate gyrus, insula, and parahippocampal gyrus (due to surface extraction artifacts caused by abrupt SP label termination leading to local over- or underestimation of surface geometry to avoid propagating errors in measurements), resulting in 17 bilateral regions. Nevertheless, while our parcellation scheme was adapted for fetal anatomy (44), it is ultimately based on adult cortical landmarks (96), which may not perfectly correspond to fetal topography, especially before sulcation. This mismatch, together with potential registration inaccuracies, particularly in regions lacking clear anatomical landmarks, may affect regional specificity. Encouragingly, the presence of systematic thickness differences and distinct developmental trajectories between neighboring regions provides indirect support that the parcellation captures meaningful spatial variation despite limited spatial precision.

Extraction of Whole-Brain and Regional SP Characteristics.

All measurements were computed in native space using registration parameters from the SP segmentation step. Whole-brain volume was the sum of CP, SP, and “other” supratentorial labels multiplied by voxel volume. SP surface area was the total area of triangles on the bilateral outer SP surface, excluding the cingular pole. SP thickness was the Euclidean distance between corresponding inner and outer surface vertices, summarized as the median across hemispheres (excluding cingular pole). SP depth, measuring the cortical folding depth as amplitude of undulations of the outer SP surface, was computed similarly to sulcal depth (adapted from ref. 102 using MIRTK (v2.0) as the median displacement along surface normals from an inflated outer SP surface, rescaled so each subject’s minimum depth was zero across hemispheres (excluding cingular pole). For each region, SP volume was calculated as the product of voxel count and voxel volume, surface area was obtained by summing triangle areas, and thickness and depth were defined as the median of vertex-wise measurements within the region.

Statistical Analysis.

Globally, our analyses first validated SP measurements and corrected for confounders (GA, residual brain volume, and regional SP depth), then assessed regional variability in SP thickness and interhemispheric asymmetry across SP characteristics. We next examined developmental changes in SP thickness, surface area, and volume from late 2nd to early 3rd trimester to identify the most developing regions and changes in asymmetry. Finally, we clustered regional growth trajectories to identify shared developmental patterns and evaluated whether lateralized features emerged within these profiles. Code and depersonalized derived data generated during analysis are available in a publicly accessible repository (103).

Regional Variation of SP Characteristics and Hemispheric Asymmetries.

To validate the extraction of SP characteristics, we first performed ANCOVA to assess effects of GA, sex, and residual brain volume (corrected for GA) on whole-brain SP thickness, surface area, and volume. Interaction terms (GA × sex, residual volume × sex) were tested. Relationships were evaluated with F-statistics and associated P-values, and partial eta squared (η²) to report effect sizes. Strong GA effects prompted post hoc visualization with linear or quadratic models chosen based on Akaike Information Criterion (AIC). Parameters were estimated with outlier-robust RANSAC regression (104), with fixed seed (42) and minimum 75% sample inclusion. Note, this strategy was applied consistently across all analyses to residualize measures for covariate corrections (including the residual brain volume mentioned above). Because regional SP thickness could be confounded by geometric variation of outer SP surface undulations (for example, SP thickness tends to be thinner at the bottom of the SP grooves; we refer interested readers to SI Appendix, Fig. S11 for analysis of SP thickness differences between sulcal and gyral regions in a central sulcus–adjacent representative area), a second ANCOVA tested thickness dependence on regional GA and residual brain volume as between-subject and SP depth (corrected for GA and residual brain volume) as within-region, within-hemisphere covariates. Significant within-subject region × depth interaction (SI Appendix, Table S2) indicated that correcting for SP depth was necessary to isolate true regional thickness differences. Regional thickness values were thus precorrected in all analyses accordingly. Finally, we modeled regional variability in SP characteristics with GA and residual brain volume as between-subject covariates, and region and hemisphere as within-subject covariates to account for the nested data structure.

To isolate regional variation, metrics were then residualized for key covariates: GA and residual brain volume, and also for regional SP depth in case of thickness. Regional variation in SP volume and surface area was not interpreted due to parcellation dependence. Absolute left–right differences were assessed using paired t tests with Benjamini–Hochberg FDR correction across regions for each SP characteristic. Additionally, normalized asymmetry indices (AI) were calculated as AI = [(right – left) × 100]/[0.5 × (right + left)] (84, 105, 106), with positive values indicating rightward asymmetry. Significant lateralization was determined by one-sample t-tests against zero with FDR correction.

Evolution of Regional SP Characteristics between the Late 2nd and Early 3rd Trimesters.

Subjects were grouped into late 2nd and early 3rd trimester groups based on GA at scan. Within each group, mean regional values were calculated after residualizing for GA and residual brain volume (and for thickness, also regional SP depth). The percent change in group means between the two periods was computed to normalize for baseline size, allowing cross-regional comparisons of SP surface area and volume. The same approach was applied to SP thickness for comparability, although absolute changes in thickness were also examined given their distinct interpretability. 95% CI were derived from SE based on group SD and sample sizes, assuming a normal distribution and a critical Z-value of 1.96. Developmental differences in relative growth between homotopic regions were assessed using Z-scores from percent change differences, with pooled SE calculated from the CIs. Asymmetry indices (AI) were computed within each group and tested for significance using one-sample t tests with FDR correction. Changes in AI between groups were evaluated using Z-tests on differences in mean AI, with pooled SE, to identify regions showing age-dependent shifts in lateralization.

Clustering to Identify Shared Developmental Dynamics.

Regional SP values were corrected for residual brain volume (and for thickness, also for regional SP depth), then modeled against GA using linear or quadratic models selected by AIC within the RANSAC framework, balancing sensitivity to nonlinearities with resistance to outliers during this period of rapid growth prior to SP dissolution (~34 to 35 wGA) (25). More flexible models (e.g., splines) were avoided due to noise sensitivity. Continuous growth trajectories were sampled every 0.1 wk between 22 to 32 wGA to produce smooth developmental profiles. Instantaneous growth rates were estimated by finite differences, normalized by baseline value to express percent growth per week, allowing size-independent comparison across regions that emphasized developmental trajectory shapes rather than absolute magnitudes (this was also performed in case of thickness for consistency) (SI Appendix, Fig. S12). Profiles were then clustered using hierarchical Ward’s method, which preserves nested structure and is robust across MRI-derived features (107) to group regions with similar developmental dynamics, with the final number of clusters selected using silhouette scores and visual inspection of dendrograms (SI Appendix, Fig. S4). Resulting clusters showed relatively high internal consistency based on silhouette scores and low within-cluster variance (though this similarity might be influenced by the choice of the modeling strategy). Cluster-averaged profiles were used to characterize dominant patterns, calculating peak growth rate, its corresponding GA, the AUC as a measure of cumulative growth, and growth density, defined as the SD of the smoothed relative growth rate curve to capture the temporal concentration of growth. Bilateral symmetry in clustering was assessed by calculating the proportion of homotopic region pairs assigned to the same cluster, normalized by the number of unique regions per cluster, with higher values indicating greater homotopic consistency in developmental trajectories.

Supplementary Material

Appendix 01 (PDF)

Data, Materials, and Software Availability

Image processing code is available on GitHub (https://github.com/FNNDSC) (108). Manual corrections and some visualizations used Freeview (3.0). Other processing used publicly available toolboxes detailed in Materials and Methods. Analyses were done mainly in Python (3.11.2) and RStudio (4.4.3). Analysis code and anonymized derived data (including cohort characteristics and quantitative whole-brain and regional measures) are publicly available on GitHub (https://github.com/AdaGondova/Gondova_2026) (103). Manual corrections and some visualizations used Freeview (3.0). Raw fetal MRI data are not publicly available due to institutional review board and data use agreement restrictions. Anonymized MRI data may be shared with qualified researchers upon reasonable request and completion of a data transfer or data use agreement. Anonymized data, including MRI scans and derivatives, will be shared upon reasonable request and signing a data usage agreement.