Insights into temporal and spatial dynamics of short association fiber formation in the human fetal brain Authors: Bo Li, Željka Krsnik, Lana Pierotich, Ivica Kostović, Simon K. Warfield, P. Ellen Grant, Davood Karimi Published: 2026 ABSTRACT Short association fibers (SAFs) form the local scaffold of cortical connectivity and support early functional specialization. However, their development before birth remains largely unknown. Using fetal diffusion magnetic resonance imaging from the Developing Human Connectome Project and histology from the Zagreb Collection of human brains, we present an in utero reconstruction of SAF pathways in the human brain. We tracked their volumetric and microstructural developmental trajectories in 243 fetuses spanning a critical period of connectome formation. We found that SAFs emerge before sulcal folding, initially as flat, loosely arranged pathways along the subplate–white matter interface, and later reorganize into coherent U-shaped bundles. Their maturation followed a sensorimotor-to-association gradient, paralleling cortical development. Nonlinear, bundle-specific trajectories captured multiphasic maturation, with subplate dynamics preceding increases in axonal coherence and early myelination. These findings reveal how the brain’s local wiring is established during prenatal development and provide insight into the origins of individual variability. INTRODUCTION Short association fibers (SAFs) are a dense network of curved axonal pathways that interconnect neighboring gyri within the same cerebral hemisphere (1, 2). They form the foundational scaffold of local cortical communication, supporting regionally confined integration crucial for cortical specialization, plasticity, and developmental tuning (3, 4). Though SAFs account for a substantial portion of the white matter (WM) volume and likely underpin key aspects of early circuit formation, they remain understudied due to their small size, high individual variability (5), and limited visibility in diffusion magnetic resonance imaging (MRI) reconstructions. These challenges are magnified in the fetal brain, where motion, low resolution, and ongoing developmental processes (e.g., cellular migration, transient zones, and dynamic tissue properties) further obscure their detection. Charting the prenatal emergence of this local wiring architecture offers crucial insight into the developmental foundations of circuit specialization and future brain function. Human fetal brain development is characterized by rapid, regionally differentiated changes in tissue composition (histogenesis), progressive establishment of corticocortical connectivity, and a developing cortex-WM interface shaped by structural and molecular changes such as subplate dynamics, laminar differentiation, and gyrification (6). Of particular interest is the distal (superficial) compartment of the developing cerebral wall, comprising the cortical plate (CP), gyral WM, and the subplate, a transient zone interposed between them (7). The subplate is the most prominent lamina of the fetal brain and persists into the early postnatal period. It serves as a site of the earliest synaptic cortical activity and a growth zone for emerging corticocortical connections, whose large size in humans has been linked to the increased number and variability of corticocortical connections compared to other mammals (8, 9). To describe the developing fetal WM, we adopted a classical five-segment division (10): segment I, periventricular zone; segment II, sagittal strata and crossroads; segment III, centrum semiovale; segment IV, gyral WM; and segment V, intracortical fiber zone. As development progresses, the abundant extracellular matrix (ECM) in the subplate diminishes, its deeper territory is progressively occupied by the expanding centrum semiovale and future gyral WM, while cortical afferent fibers that had temporarily waited in the subplate relocate into the overlying CP (6). In addition, the fetal brain undergoes a progressive shift in fiber-architectonic arrangement during the second half of gestation (20 to 40 weeks), from predominantly radial and tangential organizations, to increasingly multidirectional organizations characteristic of emerging corticocortical connectivity (11–13). During this time, projection pathways are largely in place, while long-range commissural and associative pathways, along with SAFs, develop in a temporally sequential but substantially overlapping manner, though all fiber systems continue to mature postnatally (14, 15). This sequence, from deep long-range systems to progressively refined superficial local connections, reflects the hierarchical assembly of structural connectivity in the developing fetal brain. However, knowledge of the developing SAFs, major components of the local circuits, remains limited, as they are difficult to trace and visualize in the human fetal brain. Even experimental animal models have been largely unsuccessful in demonstrating the trajectories of developing fiber pathways, with the notable exception of studies in fetal monkeys that used invasive prenatal access and autoradiographic tracers to reveal select cortical connections (16, 17). In vivo visualization using tractography offers a noninvasive alternative for investigating this important fiber system. However, existing fetal brain studies have focused exclusively on long-range pathways and reconstructed only a small number of major tracts (18–21). SAFs, by contrast, are short, spatially heterogeneous, and embedded within the fluid-rich subplate zone near the rapidly folding CP. Their small size, coupled with the small physical dimensions of the fetal brain itself, makes these fibers especially difficult to resolve using existing tractography methods. Moreover, most existing WM atlases and fiber bundle parcellation strategies are designed for large, major bundles and depend on adult-derived regions of interest (ROIs) and fixed-depth sampling approaches (22–24), which do not generalize to the rapidly changing organization of the developing fetal brain. As a result, fundamental questions remain unresolved: When do SAFs first exhibit a coherent structure? How do they develop in size, orientation, and microstructural organization throughout gestation? To address these unresolved questions, we developed and applied a computational framework tailored to SAF reconstruction in the fetal brain (see schematic in Materials and Methods). This histology-informed, gestational age–adaptive pipeline incorporates developmentally matched tissue priors, optimized tractography for SAF reconstruction in fetal diffusion MRI, and gestational age– and location-matched postmortem histological comparison to support anatomical fidelity. We applied this pipeline to multishell fetal diffusion MRI from the Developing Human Connectome Project (dHCP) (25), comprising 243 normally developing fetuses between 22 and 38 gestational weeks (GW). The postmortem histological specimens were obtained from the Zagreb Collection of human brains (26), a well-established repository of developing human brain tissue. The framework enabled robust identification of more than 250 reproducible SAF bundles confined to the superficial compartment, while excluding intermediate and deep zones such as WM segments I to III. We characterized the spatiotemporal development of SAF pathways using volume and free water–eliminated diffusion tensor imaging (DTI) metrics. This study provides the first whole-brain, gyral-level characterization of SAF development in utero, establishing a spatiotemporal reference for the maturation of short-range connectivity during a critical period of cortical assembly. These findings not only fill a long-standing gap in our understanding of the fetal connectome but also lay the foundation for normative baselines essential for identifying atypical neurodevelopment. By publicly releasing the framework and quality-controlled results, this work offers a scalable framework to map human brain connectivity before birth, advancing both basic neuroscience and clinically translational research beyond existing postnatal preterm models and enabling in vivo investigation of the prenatal origins of neurodevelopmental vulnerability. RESULTS Successful reconstruction of SAFs in utero To systematically chart the spatiotemporal development of SAFs across the fetal cortex, we analyzed high-resolution, multishell fetal diffusion MRI from the dHCP. This dataset provides research-grade in utero diffusion imaging quality that enables SAF reconstruction beyond what is currently feasible with typical clinical fetal imaging. We developed a fully automatic, fetal-optimized analytical pipeline tailored to the challenges of SAF reconstruction, informed by developmental neuroanatomy. Whole-brain fiber pathways were reconstructed using anatomically constrained probabilistic tractography (27), seeded from a curated CP-subplate interface, which enabled pathway continuity through regions where CP oversegmentation would otherwise obstruct propagation, such as at the gyral crown or beneath the insular cortex. To accommodate the diverse geometries of developing fibers and the rapidly changing microstructure across gestation, we implemented an ensemble tracking strategy with 10 tractography configurations varying in angular threshold, fiber orientation distribution (FOD) amplitude cutoff, and curvature sensitivity. The resulting ensemble included 1 million streamlines per individual. To isolate SAFs, we applied an age-adaptive superficial compartment mask that included the subplate and emerging gyral WM but excluded streamlines traversing or terminating in deeper compartments, including the intermediate zone (long-range association fibers), periventricular zone (commissural fibers), and subcortical or brainstem regions (projection fibers). Anatomical definition of this mask was guided by the laminar organization of the cerebral wall (7, 10, 28) and the CRL fetal brain atlas (29). For fetuses under 32 weeks, the subplate was delineated directly from the atlas. For later gestation, when the subplate becomes indistinct, we refined the mask through varying degrees of WM erosion followed by an iterative, inside-out “peeling” procedure to conservatively restrict streamlines to the superficial compartment. For example, beneath the insular cortex, SAFs were constrained to remain above the external capsule, the outermost deep WM structure (6). The resulting tractograms achieved high topological coverage and enabled in utero visualization of emerging SAFs during early stages of development. Unlike adult SAFs, which are often referred to as U-fibers, fetal SAFs did not initially exhibit this arcuate configuration (Fig. 1A). Instead, early pathways appeared as flat structures that gradually reshaped into curved trajectories, paralleling sulcal deepening over time. The morphological changes of the SAF system closely followed cortical gyrification, progressing from the central sulcus toward the poles and mirroring the known sensorimotor-to-peripheral spatiotemporal gradient of cortical folding (30, 31). For instance, bundles surrounding the central sulcus exhibited early curvature, followed by those in the inferior frontal (including Broca’s area) and superior temporal gyri (including Heschl’s gyrus), and later by bundles in higher-order associative regions such as the inferior temporal and superior/middle frontal gyri. By 35 GW, widespread U-fibers and coherently oriented SAF bundles were observed (Fig. 1C), similar to prior observations at term age (42 GW) (11). Progression of SAFs formation across the human cortex, revealed by in utero diffusion MRI. (A) Whole-brain tractograms at five key gestational stages (23, 26, 29, 32, and 35 weeks) demonstrate the spatiotemporal developmental gradient of superficial pathways in the fetal brain. (B) Focusing on the precentral-postcentral connection, we illustrate a developmental transformation of arcuate SAF geometry over time. The background slice represents a medial plane and does not capture the lateral cortical surface; thus, some SAFs may appear to traverse the ventricular region due to the projection of 3D bundles onto the 2D image. (C) By 35 GW, widespread SAFs with canonical U-shaped geometry are visible throughout the cortex, resembling adult-like short-range connectivity. Shown are 3D renderings from lateral, anterior, inferior, and medial (left hemisphere) views. (D) Streamlines were grouped into candidate bundles based on their endpoints, each linking a unique pair of adjacent gyri or distinct regions within the same gyrus. The illustrative results in this figure were performed on the dHCP fetal template (94) using the proposed fetal-specific analysis pipeline. Fiber orientation in (A) and (B) follows standard diffusion tensor imaging color coding: red = mediolateral, green = anteroposterior, and blue = superoinferior. A particularly dynamic transformation in the geometry of SAF bundles was observed, illustrated here by the precentral-postcentral connection (Fig. 1B). To visualize this reorganization, three-dimensional (3D) fiber bundles are overlaid on 2D sagittal T2w MRI slices, allowing the bundle geometry to be interpreted together with the developing fetal cerebral wall, including WM layering and cortical folding across gestation. At 26 GW, this intergyri bundle appears as a loosely organized structure with two to three dominant orientations, involving radial [likely reflecting neuronal migration pathways (32)] and tangential trajectories. Between 29 and 32 GW, these early, partially intersecting pathways progressively reorganized into more coherent, compact, and arcuate forms that delineate the deepening sulci along the subplate-WM interface. The presence of multiple orientations and diffuse geometry at early stages is possibly related to the underlying complexity of the subplate zone, characterized by a water-rich ECM, various migrating and developing cells, and a plexiform meshwork of growing axons with variably oriented neuronal processes (8). The observed orientation clusters may represent a macroscopic signature of this temporally dynamic and structurally complex axonal environment within the developing subplate. This flat-to-arcuate transition between 29 and 32 weeks occurred within the broader framework of strata-like ingrowth of corticocortical fibers, reflecting a progression toward spatially confined and directionally coherent bundles. By 35 GW, these pathways increasingly adopted the characteristic U-shape commonly observed in the adult brain. Histology-MRI correspondence To establish the biological fidelity of the in utero reconstructed SAFs, we compared them with postmortem histological sections from human fetal brains. These comparisons also informed the interpretation of SAF architecture within its laminar context, using histological markers. Acetylcholinesterase (AChE) histochemistry reveals key features of the developing laminar organization of the fetal cerebral wall, with intense reactivity highlighting the subplate and CP. These histological landmarks informed the design of the superficial compartment mask, defined on T2-weighted MRI, to isolate SAFs from whole-brain pathways (see Materials and Methods for details). For direct comparison of fiber architecture, we focused on later gestational stages (35 to 40 GW), where myelin- and subplate-associated markers are available through myelin basic protein (MBP) and chondroitin sulfate (CS-56) staining. In utero diffusion MRI results at 35 GW were selected to match the anatomical levels of the MBP and CS-56 staining, ensuring consistency in regional anatomy and developmental stage. In the central region at 35 GW, diffusion MRI consistently revealed SAF pathways bridging the precentral and postcentral gyri along the wall of the central sulcus (Fig. 2, A and B). These reconstructed streamlines showed precise spatial correspondence with MBP-stained SAFs located immediately beneath cortical layer VI. On MBP-stained sections, these SAFs were anatomically distinct from both the deeper extension of the immature centrum semiovale, which contains long associative pathways, and the major projection pathways streaming from the internal capsule, confirming their confinement to the superficial compartment. Distinct classes of cerebral pathways develop in parallel, highlighting the transient multilaminar organization of fetal cortical connectivity, which evolves dynamically throughout gestation. Although these fibers had not yet developed mature myelin, the associative parietal somatosensory cortex exhibited early myelination of SAFs by 35 GW, supporting their relatively advanced maturation compared with other cortical regions (33). Histological correspondence of reconstructed SAF systems. Side-by-side comparisons of in utero MRI findings (A, C, and F) and postmortem histological sections (B, D, E, and G) at matched anatomical locations illustrate the fibrillar geometry of SAFs and their position relative to adjacent cortical compartments. [(A) and (B)] Central cortex, 35 GW: Reconstructed arcuate SAFs along the wall of the central sulcus [(A), white arrow] correspond to MBP-stained SAFs located immediately beneath cortical layer VI [(A), black arrow]. (B) “1” indicates SAFs, “2” denotes the immature centrum semiovale containing long associative pathways, and “3” indicates major projection pathways from the internal capsule. [(C) and (D)] Temporal region, 40 GW: MBP staining shows a tightly bundled SAF layer [(D), blue bracket; ∼407 μm] located within developing gyral white matter beneath the SPrm [(D), gray bracket; ∼501 μm] and cortical layers I to VI [(D), red bracket; ∼1.39 mm]; the corresponding SAFs are identified from in utero MRI in (C) (white arrow). [(E) to (G)] Frontoparietal region, 40 GW: CS-56 staining delineates the SPrm as a pale band [(G), asterisks], narrowest at the sulcal fundi [(G), green bracket; ∼320 μm]; SAF location is shown in MRI in (F) (white arrowhead) and in histology in (G) (black arrowhead). W, weeks of gestation. In the temporal lobe at 40 GW, MBP staining identified a clearly delineated SAF layer located predominantly in the newly developing gyral WM, beneath the subplate remanent (SPrm) and cortical layers I to VI (Fig. 2, C and D). In this specimen, the SAF layer measured ∼407 μm in thickness, the SPrm measured ∼501 μm, and the cortical layers measured ∼1.39 mm. Although histological tissue processing can induce measurable deformation and shrinkage, the thickness of these layers highlights the need for sufficient spatial resolution and biological specificity. The SPrm contained loosely arranged plexiform fibers and remained clearly distinguishable from the dense U-shaped SAF bundles, particularly at the sulcal fundi. The corresponding SAFs were consistently identified from in utero MRI at 35 GW, indicating that the framework could capture these pathways even before the near-term stage represented by the histological specimen. Oligodendrocytes were visible on the MBP-stained section as dark dots. Fibers and dots together indicate that there is a myelination process going on in close proximity to the subplate. Additional CS-56–stained sections further delineated the SPrm as a pale extracellular matrix–rich band in the frontoparietal region near term (Fig. 2, E to G). Regional variation in SPrm thickness was evident, measuring ∼320 μm at the sulcal fundi, ∼346 to 370 μm along the gyral walls, and ∼410 μm at the gyral crowns. These laminar profiles reflect ongoing subplate dissolution and cortical compartmentalization during late gestation (6). In parallel with this reduction of the subplate, the gyral WM expanded and progressively occupied the former subplate position in the proximal portions of the gyri. Within this developmental context, the principal location of SAFs appeared to shift from within the subplate in earlier stages to the underlying newly formed gyral WM near term. Together, these histological observations provide microstructural context for interpreting SAF development and support the anatomical fidelity of the proposed in utero MRI-based analytical framework, including its ability to localize SAF systems relative to cortical layers, subplate-related compartments, and deeper long-range fiber pathways. In the frontal lobe, a coronal section at 40 GW revealed multiple SAF systems spanning diverse cortical territories, together with other major axonal pathways visible on MBP-stained histology, including prominent projection fibers passing through the internal and external capsules (Fig. 3). Higher-magnification histological views identified SAFs immediately beneath cortical layer VI. Myelin-stained sections and in utero MRI jointly delineated several corresponding SAF connections between the middle frontal and the opercular portions of the inferior frontal gyri (Fig. 3B), between the superior and middle frontal gyri (Fig. 3E), and between the supplementary motor area and the mid-cingulate cortex (Fig. 3H). These pathways traversed gyri of varying morphological complexity, including both middle and late folding gyri (table S1), and extended across both lateral and medial surfaces. It is additionally important to see from the myelin stains that the frontal associative regions mature not much behind the somatosensory cortex. These frontal comparisons further support the anatomical validity of in utero SAF reconstructions. Although histological correspondence was available only for near-term specimens, the MRI-based reconstructions extend across a substantially earlier developmental window, enabling investigation of SAF organization throughout the extremely preterm, very preterm, and preterm stages. Histological correspondence of frontal SAF systems. MBP-stained coronal sections from the frontal region of a 40-GW human brain and corresponding in utero MRI-derived SAF reconstructions illustrate frontal SAF geometry across multiple cortical territories. (A to C) An overview histological section show multiple axonal fiber systems, including prominent projection fibers passing through the internal capsule [dashed line in (G)] and external capsule [arrowhead in (G)]. (G to I) Higher-magnification views of the boxed regions in (A) to (C) highlight SAFs located immediately beneath cortical layer VI (black arrows). (D to F) Corresponding SAF pathways reconstructed from the dHCP fetal template at 35 GW show tight arcuate pathways (white arrows) whose geometry and location align with those observed histologically. These include connections between the middle frontal gyrus and the opercular portion of the inferior frontal gyrus (D), between the superior and middle frontal gyri (E), and between the supplementary motor area and the mid-cingulate cortex (F). W, weeks of gestation. Bundle-level reproducibility reveals spatiotemporal emergence patterns Building on the successful in utero reconstruction of SAFs and their confirmed anatomical validity against postmortem histology, we next examined how these pathways emerge and morphologically change over gestation. Specifically, we assessed the reproducibility of individual SAF bundles across fetuses within each age group to identify robustly expressed pathways and to assess their consistency across age groups, resulting in approximately 250 reproducible bundles per age group. Beyond filtering out spurious streamlines, bundle-wise reproducibility provided a quantitative marker of population-level maturation, indicating when and where coherent SAF geometries begin to consistently manifest across individuals. We defined the temporal onset of each bundle as the earliest gestational age at which it became consistently reproducible across all later groups. Using the 36-GW reproducible set as a reference, we traced each bundle backward to determine when this sustained group-level presence began. To explore regional differences in maturation, bundles were grouped by connected lobes (e.g., frontal-frontal and parietal-temporal). To examine its associations with cortical developmental timing, bundles were additionally grouped by the folding stage of the connected gyri (early, mid, or late) based on established histological timelines (30). For this analysis, each bundle was assigned conservatively according to the later-developing gyrus of the connected pair, so that bundles linking gyri from different folding stages were categorized by the more developmentally delayed component. The distribution of onset ages revealed pronounced spatial heterogeneity (Fig. 4A), with a clear central-to-peripheral maturation gradient. Bundles in parietal and parieto-frontal regions, particularly those associated with the sensorimotor and perisylvian cortices, emerged earliest and remained reproducible throughout the third trimester (Fig. 4C). In contrast, bundles in more peripheral regions, including the prefrontal and occipital cortices, emerged later. This spatiotemporal pattern echoes postmortem DTI findings (11) and parallels the known sequence of cortical maturation (34, 35). Regionally, early-developing bundles predominantly involved primary sensory and motor areas (e.g., postcentral, supramarginal, and precentral gyri), followed by a progressive anterior expansion from premotor areas to the superior, inferior, and middle frontal gyri, eventually reaching the frontal pole. Posteriorly, development extended from medial occipital areas to the lateral occipital and inferior temporal gyri. This spatial gradient also mirrored the sequence of cortical folding: Bundles connecting early-visible gyri tended to exhibit directional coherence and reproducibility at younger gestational ages than those involving mid- or late-developing gyri (Fig. 4, C and D). Categorical analyses based on lobar location and folding stage further suggested that bundles connecting anatomically or developmentally similar regions, such as within-lobe or within-stage pairs, generally emerged earlier than those bridging more distinct regions or developmental stages. While the inclusion of intragyri bundles may have contributed to earlier reproducibility within some categories, this alone did not explain the overall pattern. The parieto-frontal group was a notable exception, emerging early despite spanning distinct lobes, consistent with the advanced maturation of the frontoparietal system. Spatial patterns and hemispheric symmetry of SAF bundle emergence. (A) Onset of reproducibility grouped by the lobe of the cortical gyri connected by each bundle, with left and right hemisphere bundles considered separately. Violin plots show the distribution of bundle-specific onset ages within each lobar group; central dots indicate medians and bars denote the interquartile range. (B) Left-right comparison of onset timing among 128 matched bundle pairs. The Bland-Altman plot shows the difference in onset age between left and right bundles against their average; the solid line indicates the mean difference and the dashed lines denote the 95% limits of agreement. (C) Onset of reproducibility grouped by the major folding stage of the connected gyri (early, mid, or late; table S1) (30). (D) Finer breakdown of onset timing across all six gyral pair combinations (e.g., early-mid and mid-late). We also assessed hemispheric asymmetry in the onset timing of reproducible bundles by comparing 128 matched left-right bundle pairs present in both hemispheres. The overall difference was not statistically significant (paired t test, P = 0.25), indicating no evidence of systematic hemispheric asymmetry in the anatomical emergence of SAF bundles. Bland-Altman analysis likewise showed the absence of a systematic trend, with most bundle pairs clustering near the zero-difference line and falling within the 95% limits of agreement (Fig. 4B). A small number of bundle pairs lay outside these limits and showed relatively large left-right differences in onset age. Given the conservative hierarchical criteria used to define reproducible emergence across age groups, such outliers likely reflect delayed confirmation in one hemisphere due to sampling variability or signal dropout in specific age groups. These factors can hinder group-level consensus even when the bundle is anatomically present in individual fetuses, making methodological delay a more likely explanation than true biological asymmetry. Trajectories of structural and microstructural maturation in individual bundles Having identified the spatiotemporal emergence gradient of SAF bundles across the fetal cortex, we next quantified the developmental trajectories of individual bundles, focusing on changes in volume and free water–eliminated DTI metrics [fractional anisotropy (FA) and mean diffusivity (MD)] to capture both macroscopic growth and microstructural maturation. All measurements were computed in each individual’s native space. Volumes were estimated from native bundle masks defined by thresholding the voxel-wise streamline density map at the 5th percentile, while microstructural measurements were sampled using group-level consensus masks warped to the individual brain. The bundle exemplars shown in Fig. 5 were extracted fully automatically, without manual editing. Developmental trajectories of representative SAF bundles in the frontal lobe. Six bundles are shown: SMA-MidCng (supplementary motor area to middle cingulate gyrus), PreC-Pstcent (precentral to postcentral gyrus), SupF-MidF (superior to middle frontal gyrus), OpIF-OpIF (within the opercular inferior frontal gyrus), TriIFG-TriIFG (within the triangular inferior frontal gyrus), and PreC-SupF (precentral to superior frontal gyrus). Representative bundle morphologies are shown at 29, 32, and 35 GW. In the developmental trajectories, left-hemisphere measurements are shown in orange and right-hemisphere measurements are shown in green; both hemispheres were jointly modeled. Bundle-specific measurements of volume (top; mm3), free water–eliminated fractional anisotropy (middle; FA), and mean diffusivity (bottom; MD, ×10−3 mm2/s) are plotted across gestational age. A 3D rendering of the entire fiber bundle at 35 GW is overlaid on a 2D sagittal MRI slice (left view) to show its overall shape and anatomical location, using the same visualization approach as in Fig. 1B. FA and MD values represent the median of nonzero values within each bundle mask. Volume was fit with a linear model, whereas DTI measures were fit with either linear or quadratic regression based on adjusted R2. Shaded regions denote the 95% confidence intervals of the fitted trajectories, estimated by nonparametric bootstrapping (1000 iterations). Figure 5 presents results from six frontal lobe bundles: precentral-postcentral (PreC-Pstcent), superior-middle frontal (SupF-MidF), precentral-superior frontal (PreC-SupF), supplementary motor area-middle cingulate (SMA-MidCng), opercular inferior frontal-opercular inferior frontal (OpIF-OpIF), and triangular inferior frontal-triangular inferior frontal (TriIFG-TriIFG). These bundles sample both intergyral and intragyral frontal SAF systems across regions with distinct developmental timing and morphological complexity. Volumetric expansion was observed across all bundles, though the rate and pattern varied, likely reflecting variations in gyral size, intergyral interface extent, and developmental timing. While most bundles exhibited near-linear growth within the studied age range, several bundles showed localized volume surges around 30 GW, such as the PreC-SupF connection. These fluctuations may reflect regional dynamics in subplate expansion and dissolution. Representative morphologies across gestation further showed that, rather than forming the static U-shaped configuration commonly seen in the mature brain, fetal SAF bundles initially appeared as relatively flat structures with several dominant orientations and gradually reorganized into more tightly bundled, directionally coherent, arcuate pathways as sulci emerged and deepened. Microstructural maturation followed a more complex, nonlinear trajectory. Free water–eliminated FA decreased in early gestation, reached a turning point (minimum), and then increased, while MD showed the opposite pattern. The turning points of MD (peaks) were earlier than those of FA, mostly between 26 and 30 GW, likely reflecting maximal subplate expansion, particularly in the frontal associative areas. FA turning points were observed later (after 31 weeks) or, in some cases, were not yet reached within the studied age range. Notably, regional variation in microstructural maturation was evident even within the frontal lobe. For example, the PreC-Pstcent bundle showed an earlier FA turning point at 31.5 weeks, compared to later turning points in PreC-SupF (36.6 GW) and SMA-MidCng (32.6 GW), while SupF-MidF exhibited an even later shift that was extrapolated beyond the study range at 39.7 weeks. These diverse trajectories highlight how local connectivity pathways follow distinct maturational schedules even within the same lobe. Such bundle-specific differences likely reflect underlying biological processes. Myelination and axonal organization progress along regionally asynchronous timelines (36), while the subplate undergoes dynamic changes in the amount and composition of ECM and changes in its fiber architectonics (6). Our findings align with histological and imaging literature suggesting that simpler laminar architecture, such as limbic areas, tend to follow simpler growth trajectories, whereas polysensory and high-order association cortices, characterized by greater cytoarchitectonic complexity, exhibit more complex, protracted developmental profiles (37). Complementary results for additional bundles are presented in the Supplementary Materials. These include SAFs in the temporal (SupTemp-MidTemp, MidTemp-InfTemp, Ang-MidTemp, and Fusiform-InfTemp), parietal (Precuneus-PCL, Pstcent-SPL, SPL-IPL, and SPL-Precuneus), occipital (SupOcc-MidOcc, MidOcc-InfOcc, Calc-Ling, and Ling-Fusiform), and insular cortices (OpIF-Ins, OrbIF-Ins, and RolOper-Ins). Full bundle names and abbreviations are listed in table S1. Together, these analyses provide a cortex-wide characterization of SAF maturation in utero across both inter- and intragyral scales, revealing regionally heterogeneous growth and highlighting the sensitivity of DTI metrics to the underlying developmental processes. Regional heterochrony in SAF microstructural developmental patterns Extending from bundle-specific developmental trajectories, we next examined the timing of key microstructural transitions to uncover broader organizational principles of SAF maturation across the cortex. Specifically, we identified turning points in free water–eliminated FA and MD curves, defined as local extrema in quadratic fits, as temporal landmarks where the direction of tissue property change reverses, indicating a shift in the prevailing biological mechanisms driving maturation. Across the cortex, MD and FA exhibited complementary temporal dynamics (Fig. 6). MD turning points consistently preceded those of FA, with MD peaks occurring between 24 and 29 weeks and extending to 31 weeks in the frontal lobe. These early MD peaks likely coincide with transient subplate expansion, a phase marked by loosely arranged axons and increased ECM. The subsequent decline in MD suggests a progressive restriction of water diffusion as axonal organization increases and as the subplate dissolves. FA, in turn, reached its minimum later, generally after 31 weeks, with subsequent increases likely driven by axonal packing, laminar differentiation of the cerebral wall, and early myelination. This temporal offset between MD and FA suggests that each metric captures distinct yet interrelated aspects of SAF development, with MD being more sensitive to early permissive isotropic environment and FA to later-phase fiber organization. Lobar distribution of SAF microstructural turning ages. Gestational age (weeks) at the turning points of bundle-specific trajectories of free water–eliminated fractional anisotropy [FWE-FA (A)] and mean diffusivity [FWE-MD (B)], grouped by anatomical lobe. Turning points were extracted from the vertex of the best-fitting quadratic model for each bundle. Bundles with monotonic trends, or with turning points estimated beyond the observed window (22 to 38 gestational weeks), were assigned a value of “>38 weeks” to indicate that no turning point was observed before 38 weeks, and were excluded from the statistical summary. Left- and right-hemisphere measurements were combined for joint modeling and counted once per regression. In both panels, violin plots show the distribution of bundle-specific turning ages within each lobar group; central dots indicate medians and bars denote the 25th and 75th percentiles, computed using only values within the observed study range (22 ≤ GA ≤ 38 weeks). The lobar groups are ordered separately in each panel according to the corresponding turning-age pattern. Colors indicate lobar groups, with the number of contributing bundles shown in parentheses. Two large-scale principles of maturation emerged from the distribution of turning points. The first reflects a spatially ordered progression that extends from the central sulcal (fissure of Rolando) toward the lateral convexity (e.g., inferior temporal gyrus) and anterior poles. For example, within the temporal lobe, the superior-middle temporal bundle reached its FA turning point by 36.8 weeks, while the middle-inferior temporal bundle followed later at 40.5 weeks (Fig. 4C). A similar pattern was observed in the occipital lobe, where the superior-middle occipital bundle (38.5 weeks) preceded the middle-inferior occipital bundle, which showed a projected FA turning point beyond term (50.5 weeks), suggesting critical microstructural changes after birth (fig. S9). The lingual-fusiform bundle, located in the lateral occipitotemporal cortex, also followed a protracted trajectory, with a turning point estimated at 42.5 weeks. These spatial patterns mirror known sequences of gyral development (30). Superimposed on this spatial gradient was a second, partially overlapping pattern that followed a functional gradient from primary sensorimotor areas to higher-order association cortices, consistent with the sensorimotor-association (or unimodal-transmodal) axis featured in human cortical organization (38, 39). Early-maturing axonal bundles included PreC-PostC (31.5 weeks) and SMA-MidCng (32.6 weeks), which are anchored in primary or supplementary motor regions, whereas bundles associated with higher-level cognitive and executive functions, such as SupF-MidF in the dorsolateral prefrontal cortex, showed later maturation (39.7 weeks). Cuneus-lingual (37.0 weeks), situated in the primary visual cortex along the calcarine sulcus, reached its turning point earlier than lingual-fusiform (42.5 weeks), which lies in the lateral visual association cortex. Within Broca’s region, we observed subtle timing variations: the OpIF-OpIF bundle (34.8 weeks), in pars opercularis [Brodmann area 44 (BA44)], exhibited an earlier estimated turning age than TriIFG-TriIFG (36.3 weeks), in pars triangularis (BA45). While modest, this distinction may reflect differential maturation related to their proposed roles in motor aspects of speech versus lexical-semantic processing (40). Other variations within Broca’s area, including cytoarchitectonic differences (BA44 dysgranular and BA45 granular) and distinct connectivity profiles (41), further scaffold this functional specialization, although specialization in this region remains a subject of ongoing investigation. Last, we compared the timing between bundle emergence and microstructural turning points to assess whether anatomical geometry and maturation unfold in synchrony (fig. S12). While early-emerging bundles often matured earlier, especially in sensorimotor areas, this temporal alignment was not uniformly observed across all bundles. These findings suggest that emergence and maturation, while partially coordinated across regions, are further shaped by region-specific developmental programs. DISCUSSION SAFs form a dense scaffold of local connectivity, facilitating early communication between neighboring gyri during cortical development. In this study, we systematically charted the anatomical and microstructural development of SAFs in the fetal brain, revealing an organized yet heterogeneous maturation process. Here, maturation of the fiber pathways refers broadly to the full spectrum of processes required for their development, including axonal growth, ingrowth, retraction, geometric organization, and myelination, as understood from anatomical and neurodevelopmental perspectives (42–44). SAFs were found to emerge early in gestation as loosely arranged, multidirectional pathways, and gradually exhibited coherent orientation and arcuate shapes that tightly delineated cortical sulci. This morphological transformation was accompanied by a progressive increase in population-level reproducibility and anatomical specificity. Quantitative analysis of more than 250 bundles across the cortex uncovered regionally distinct patterns of formation and maturation, with earlier development in sensorimotor regions and delayed trajectories in higher-order association areas. Microstructural measurements exhibited nonlinear changes over gestation, with turning points in FA and MD revealing spatial heterochrony in tissue maturation. Together, these findings demonstrate that SAF development follows a coordinated yet regionally variable timeline shaped by both spatial and functional hierarchies, providing a foundation for understanding the early emergence and formation of local cortical circuits. The time of appearance and positioning of SAFs along the radial axis of the cerebral wall in human fetuses can be understood within the broader, sequential yet overlapping maturation of distinct classes of cerebral pathways. Our findings refine this understanding by charting the in utero trajectory of SAF maturation with unprecedented spatiotemporal detail, revealing that SAFs emerge during the mid-to-late phase of cortical connectivity development, after long-range projection pathways are in place, but before intracortical circuits develop postnatally. This temporally orchestrated sequence begins with modulatory afferents from brainstem nuclei, arriving before 8 GW and targeting the marginal zone and subplate. Shortly thereafter, afferents from the basal forebrain and other subcortical telencephalic nuclei begin to populate the subplate. Following the establishment of the CP around 8 GW, thalamocortical projections undergo extensive development, with fibers penetrating the CP only much later, around 22 to 23 GW, after forming transient synapses in the subplate. The corpus callosum begins to form after 10 GW, likely arising from early layer III pyramidal neurons whose axons project contralaterally. Around this period, one of the earliest short corticocortical circuits, the reciprocal connections between the entorhinal cortex and hippocampus (archicortex), also begins to form, supporting early limbic network development before the maturation of neocortical associative systems (15). Long-range associative fibers, also originating from layer III pyramidal neurons, begin to emerge at preterm stages (subsequent to callosal formation), and exhibit both ipsilateral and contralateral collaterals. The uncinate fasciculus is among the earliest long association tracts to appear (45), and by approximately 20 GW, most major pathways become identifiable in dissection studies (46). The SAFs described in the present study, develop in parallel with these long-range systems but are more closely linked to the expansion of local cortical circuits (47). Last, intracortical fibers, including tangential connections within the cortical layers, predominantly develop after birth, as supported by postnatal histological evidence (14). The earliest stages of superficial pathway development are best interpreted within a gradual and region-dependent framework of corticocortical network assembly, rather than as a yes or no question with a single universal onset age. In this context, the estimated onset of bundle reproducibility represents a conservative population-level marker of sufficient anatomical coherence and cross-subject consistency, not the direct biological onset of axonal growth or bundle formation. Prior ex vivo diffusion MRI studies have reported emergent short-range corticocortical pathways as early as 17 postconception weeks in selected parietal and occipital regions, coursing through the subplate and intermediate zone, with broader corticocortical organization becoming more evident later in gestation (11). At the same time, diffusion-defined pathways at these early stages likely reflect a mixed developmental environment with changing tissue composition that includes various migrating and developing cells, and a plexiform meshwork of growing axons with variably oriented neuronal processes, which warrants cautious interpretation of the youngest SAF pathways (12, 13, 48–50). Within this framework, at the youngest ages, regional FWE-DTI measures were interpreted as characterizing the developmental state of the superficial compartment, whereas later ages additionally allowed clearer interpretation of coherent SAF bundle geometry. This perspective also helps explain the larger dispersion of early-age diffusion measures, because the superficial compartment is biologically more heterogeneous at these stages, with multiple developmental processes occurring simultaneously in a transient and rapidly changing environment, whereas at later ages the tissue environment is relatively more stabilized and structurally consolidated (6). The fetal-stage development of SAFs coincides with a critical transitional window of cortical network assembly, as local circuits begin to form on the scaffold of long-range connectivity. Anatomically, SAFs are confined to the superficial compartment between the CP and deeper WM segments, a region dominated during this period by the subplate. Their development, therefore, closely parallels subplate maturation, which presents a major hub of early synaptogenesis, axonal pathfinding, neuron-neuron-glia interaction, and early network coordination (51). Notably, SAF expansion accelerates between 28 and 32 GW, during the peak thickness of the subplate and a period of increased vulnerability to perinatal brain injury. These spatiotemporal patterns suggest that SAFs help establish local communication within the global brain network and may serve as in utero markers of cortical circuit maturation and early neurodevelopmental risk. Spatial heterochrony across structural and microstructural milestones emerged as a defining feature of SAF development, with anatomical establishment and underlying biological processes following regionally asynchronous timelines. Bundles in primary sensorimotor regions appeared earliest, including parietal and frontal-parietal areas, e.g., the prominent precentral-postcentral U-fiber system, which has been well described in the mature brain (23). The spatiotemporal gradient of SAF development was coordinated with the broader principles of fetal brain maturation, in which regional specialization, laminar organization, gyrification, and subplate dynamics influence and are influenced by the maturation of cortical connections (51, 52). Early structural coherence in primary sensorimotor bundles aligns with functional observations in preterm neonates: Coordinated motor and sensory bursts are detectable by EEG as early as 26 GW (53). Complementary animal studies show that spontaneous retinal activity propagates across the visual system before sensory experience begins (i.e., before eye opening), suggesting that coherent structural pathways can precede and help organize early functional activity (54). Within this study, our observation of early and robust SAF development in the parietal lobe (postcentral gyrus; superior/inferior parietal lobules; precuneus) corresponds with a marked shift in sensorimotor integration and motor coordination in fetal behavior (55). In addition, tactile input begins to be coordinated with motor output, reflecting increased engagement of the parietal and frontal networks and emerging multisensory integration, which parallels the early maturation we observed for frontal-parietal connections. This also includes the early coordination of auditory input with motor activity, likely mediated by a network involving the posterior parietal cortex, superior temporal gyrus, inferior frontal gyrus, and angular gyrus (56), and includes the emergence of resting-state networks (including auditory, somatosensory, motor, and frontoparietal) during the second and third trimesters (57, 58). Together, these converging lines suggest that early-developing local association pathways may contribute to the scaffolding of functional circuits during a critical window of network formation. The region-specific nonlinearity of FA and MD trajectories and the systematic lead of MD peaks over FA minima support a multiphasic model of SAF maturation: early phases dominated by subplate expansion with a large and hydrated extracellular space (rising MD, declining FA); a transitional phase characterized by the accumulation of loose plexiform network of fibers and postmigratory cells (declining MD); and later phases marked by subplate ECM dissolution, emergence of gyral WM, tightly packed and orderly arranged axons, and myelination that strengthen directional coherence (rising FA), with earlier timing in primary sensory and motor regions. The transitional phase characterized here, manifesting as a decline in MD before FA increase, largely corresponds to the “premyelination” stage described in earlier imaging studies, which noted restricted water diffusion at developmental stages preceding histologically confirmed myelination (59, 60). This phase is thought to be associated with increases in axonal caliber, the initial ensheathment of axons by oligodendrocytes, and, within the subplate zone, a 3D reorganization of the axonal plexus (61, 62). Although free water–eliminated DTI mitigates contamination from extracellular fluid, the resulting measures still reflect a composite change of tissue microstructure. In addition, it is worth noting that the nonlinear FA trajectory is better interpreted not as a direct readout of bundle geometry, but as reflecting the evolving diffusion-derived tissue property of the superficial compartment, of which visible fibrillar reorganization represents only one component of the broader developmental change. Placed alongside existing preterm models, our in utero findings show both shared patterns and important divergences that arise from differences in developmental coverage (a large cohort spanning 22 to 38 GW), biological milieu (in utero versus ex utero), and anatomical specificity. Prior “prenatal” SWM inferences predominantly come from neuropathological studies obtained postmortem, or very preterm neonates (<30 GW) scanned ex utero. Those studies describe spatially heterogeneous maturation, with faster FA increases in primary sensorimotor cortex than in association areas (63, 64), a pattern that parallels our regional gradients. Key differences, however, highlight why an in utero baseline is needed. First, most preterm studies report linear FA increases from approximately 28 to 32 GW onward, consistent with their sampling window, whereas our in utero coverage back to 22 GW and bundle-level analysis reveal earlier, bundle-specific nonlinear phases that precede later increases. These early phases, including an initial FA decline, would be difficult to detect in cohorts that begin after 26 GW and include few observations between 26 and 30 weeks, limiting the power to identify earlier stages of the development, particularly those occurring before cerebral pathways are established. Second, in preterm cohorts, brain injury can obscure developmental patterns due to anterograde and retrograde degeneration, and compensatory reorganization (65). Axonal pathways are still growing during the early preterm stage, making WM in preterm infants markedly different from that of term newborns (66–69). Moreover, the developmental milieu differs fundamentally: Very preterm infants mature ex utero, where altered sensory input, stress, and nutrition can influence neurodevelopmental trajectories, sometimes delaying features [e.g., sulcation (70) and widespread FA reductions (71)], and sometimes accelerating others [e.g., experience-driven FA increases in visual cortex (72)]. A third difference lies in anatomical precision. Many prenatal studies approximate SWM, WM (73–75), or the subplate (76), using fixed-depth sampling beneath the cortex or manually placed ROIs, risking inclusion of adjacent compartments, and overlooking regional and gestational dynamics of the fetal cerebral wall. In contrast, our framework leverages refined spatial grids (0.5 mm isotropic), anatomically guided tractography, bundle-specific modeling, and gestational age–adapted tissue segmentation to explicitly account for the developing interface between superficial and deeper compartments, enabling gyri-level resolved SAF mapping across the fetal cortex. Together, these features underscore the unique value of in utero diffusion MRI in establishing a normative reference for SAF development, free from prematurity-related confounds and optimized for fetal neuroanatomy. MRI-histology correspondence supports the anatomical specificity of the in utero SAF reconstructions, strengthens the interpretability of diffusion-derived measurements as highly sensitive in vivo indices of tissue changes within the appropriate laminar context, and highlights the advantages of in utero diffusion MRI, i.e., fine temporal sampling, 3D quantification and visualization, and whole-brain coverage, relative to the necessarily sparse sampling of postmortem sections. More broadly, these results illustrate the value of anatomically grounded in utero imaging for resolving early SAF organization during its emergence and provide a scaffold for future integrations with in vitro postmortem imaging and molecular datasets to probe regional vulnerability and the development of circuit specialization. Beyond establishing an in utero baseline and aligning MRI with histology, we introduce a set of methodological advances that are critically needed for tracking and visualizing connectivity in the developing human brain (77, 78). First, we operationalize an anatomically faithful definition of fetal SAFs, confined to the subplate and emerging gyral WM, using a histology-informed, “inside-out” peeling strategy and age-adaptive tissue segmentation that respects the transient subplate zone and the developing SP-WM interface. Second, we pair anatomically constrained, ensemble tractography with gyri-resolved parcellation and a hierarchical quality-control workflow (individual pruning → group normalization to weekly templates → bundle-wise reproducibility within and across age groups). This supports gyri-level mapping of short-range pathways across 22 to 38 GW, enabling bundle-wise developmental profiling and spatiotemporal modeling of both global maturation patterns and short-range structural connectivity across the fetal cortex. Third, we quantify maturation with bundle-specific macrostructural features and free water–eliminated DTI metrics, which help mitigate fluid contamination in the highly hydrated fetal brain. Despite these strengths, several limitations frame interpretation and naturally suggest next steps. (i) Developmental fate. Within the fetal window, we cannot determine the extent to which early SAFs confined to the superficial compartment eventually contribute to long-range associative systems, which are progressively displaced into deeper WM with maturation. (ii) An important limitation is that the postmortem histology and in vivo diffusion MRI were obtained from different individuals, and the available myelin staining is concentrated in later gestation. Because direct validation of in vivo tractography is not currently technically or ethically feasible in the human fetal brain, gestationally matched histology provides the closest available approximation to ground truth for interpreting laminar location and pathway geometry. Accordingly, the histological analysis is presented as qualitative anatomical correspondence rather than direct validation, and the earliest developmental stages are interpreted within a more graded framework that combines in vivo observations with prior ex vivo diffusion MRI literature. Absolute thickness comparisons should also be interpreted cautiously, because histological tissue processing can induce measurable deformation and shrinkage. (iii) Temporal sampling. Our cross-sectional design (22 to 38 GW) precludes within-subject developmental trajectories and leaves some turning points extrapolated beyond term. Extending future analyses into immediate postnatal windows would help better constrain the model-derived turning ages. (iv) Microstructural modeling. Free water–eliminated DTI improves specificity and is easy to interpret, but DTI remains a simplified model of complex microstructure. Adding multicompartment modeling and higher-b shells, where feasible, should sharpen biological attribution (e.g., axonal density versus anisotropy). (v) Group-level masking. Consensus bundle masks enhance robustness for regression yet can attenuate subtle individual variations, a trade-off inherent to large-scale, automated pipelines. Because bundles with more complex or heterogeneous shapes may be represented primarily by their spatially consistent core, the proportion of each bundle captured by the consensus mask may differ across regions and should be considered when interpreting regional comparisons. (vi) Conservative estimation. Our estimates of SAF onset timing are deliberately conservative, anchored to persistent group-level presence across all subsequent age groups. This approach strengthens robustness and prevents spurious detections, but may lag behind actual anatomical emergence. Similar delays have been noted in prior fetal MRI studies, where sulcal appearance lags behind postmortem reports (79), and initial identification occurs before consistent presence (80, 81). This conservative criterion may exaggerate left-right differences when emergence is confirmed in one hemisphere but not the other due to signal dropout or small group sizes with variable data quality, where our group-level criteria are harder to meet. Accordingly, extreme hemispheric differences (more than 10 weeks in some cases) likely reflect methodological constraints rather than genuine asymmetry. Looking ahead, several opportunities could broaden the scientific reach of this work. Region-focused analyses (e.g., perisylvian language cortex) can test for subtle asymmetries, such as volume and morphology, and specialized timelines that whole-brain or lobar summaries may dilute. Multimodal extensions pairing fetal fMRI with bundle-wise diffusion metrics would probe evolving structure-function coupling in utero. Relating bundle-wise diffusion trajectories to regional subplate dynamics would further clarify how SAF emergence and maturation are coupled to the expansion and dissolution of this transient fetal compartment. Applying the pipeline to high-risk cohorts or fetuses with high-risk genotypes will help contextualize deviations against the normative reference established here and refine early biomarkers. Last, community release of code and quality-controlled results will enable benchmarking and encourage harmonization efforts and test-retest studies. These resources may also support further improvements in individual-level tract specificity through refined fetal cortical parcellation and SAF-specific endpoint and exclusion criteria, particularly as fetal cortical labeling and transient-zone segmentation continue to improve. Integration with histology and transcriptomic datasets can link regional trajectories to underlying cellular programs, improving mechanistic interpretations of SAF maturation. MATERIALS AND METHODS We conducted the study in five main steps using fully automated methods (Fig. 7), including (i) data inventory and processing; (ii) fiber tracking and isolating superficial fibers from deep, long-range fibers; (iii) parcellation of inter- and intragyri SAF bundles; (iv) group-wise evaluations for each age group; and (v) longitudinal modeling of bundle-specific morphological and microstructural changes across gestational ages and lobe-wise analysis. The first three steps were progressively refined through manual inspections and comparison with neuroanatomy knowledge. Pipeline for spatiotemporal analysis of fetal SAF development. The framework integrates gestational age–specific preprocessing, developmentally refined tractography, and a histology-informed exclusion mask to isolate SAFs from deep, long-range fibers. Following fiber bundle parcellation, developmental modeling characterizes bundle-wise trajectories of volume and microstructure (FWE-FA and MD) across gestation. Additional analyses identify the onset of bundle reproducibility, hemispheric asymmetry, and microstructural turning points stratified by lobar group and gyral type. Together, this framework enables in utero, gyral-level, whole-brain characterization of SAF maturation from 22 to 38 gestational weeks. Datasets and MRI processing The brain specimens used for postmortem histological comparison were obtained from the Zagreb Collection of human brains (26), a well-established and diverse repository of human fetal brain tissue. The histology study was conducted according to the guidelines of the Declaration of Helsinki, and approved by the Ethics Committee of the School of Medicine University of Zagreb (protocol number Ur. Broj: 251-59-10106-23-111/158, Klasa: 641-01/23-02/01). Fetal brain MRI from the dHCP was used for the study of SAFs in utero (25). The dHCP study was approved by the UK Health Research Authority (Research Ethics Committee reference 14/LO/1169), and written informed consent was obtained from participating families before imaging and for release of anonymized data. This dataset presents the highest quality open-source data for fetal brain MRI, allowing for cutting-edge and reproducible studies by the entire scientific community. Diffusion-weighted MRI (dMRI) data were collected with a combined spin echo and field echo sequence at 2-mm isotropic resolution, using a multishell diffusion encoding that consists of 15 volumes at b = 0 s/mm2, 46 volumes at b = 400 s/mm2, and 80 volumes at b = 1000 s/mm2 (82). In addition, T2-weighted MRI was acquired from multiple uniquely oriented stacks and reconstructed into a 0.5-mm isotropic 3D volume by slice-to-volume reconstruction, which provided the structural basis for the anatomical priors used in the tractography pipeline (83). Standard preprocessing operations that have been performed on the released data include denoising, eddy current and susceptibility-induced distortion correction, and motion correction (84). Because the native fetal dMRI resolution remains limited, we linearly up-sampled the data to an isotropic voxel size of 0.5 mm to provide a finer computational grid for anatomically constrained tractography, to reduce discretization effects during streamline propagation, and to improve compatibility with the high-resolution structural priors used by the pipeline. This step was not intended to imply recovery of additional independent sub-voxel diffusion information. The MRI data and intermediate results derived from them were standardized to “right-anterior-superior” orientation. Detailed cortex parcellation and fetal brain tissue segmentation, including transient zones such as the subplate, were obtained by nonlinearly aligning the T2-weighted MRI and propagating the CRL fetal brain atlas (29, 85) of matched age to individual space (86). The classical model for estimating the local orientation of prominent fibers in each imaging voxel is DTI. Although DTI is inherently limited to representing a single fiber orientation per voxel, it remains a widely used and robust approach, particularly in situations where more complex models cannot be reliably applied. Methods such as spherical deconvolution estimate a full FOD that has the potential to resolve crossing fibers. These methods, however, require more measurements and a higher signal-to-noise ratio for reliable model fitting. Moreover, recent studies have shown that standard FOD estimation methods are not optimal for early brain development, where the tissue microstructure is vastly different from adult brains (87). In this work, we adopted a practical strategy for choosing a suitable model of local fiber orientations by computing and assessing four different representations derived from the dMRI data. (i) Single-shell single-tissue constrained spherical deconvolution (CSD) using the b=0 and b=1000 s/mm2 data (88); (ii) multishell multitissue CSD using the data in all three shells (89); (iii) DTI estimated with a weighted linear least squares method; and (iv) sharpened diffusion orientation distribution function (dODF) computed from the diffusion tensor (90). We computed tractography using these four input models, followed by visual inspection and qualitative comparison of the resulting whole-brain tractograms and the extracted superficial bundles. On the basis of this assessment, we adopted the DTI-derived dODF for subsequent analysis, as it provided broad topology-level coverage and continuous arcuate trajectories along the gray matter–white matter interface (GMWMI) compared with the alternatives. Last, we computed diffusion tensor scalar measures, i.e., FA and MD, from both the standard DTI and free water–eliminated DTI (91). Tractography and SAF systems To seed and constrain fiber tracking using anatomical priors, we computed five-tissue-type (5TT) segmentation maps from the tissue segmentation provided in the official dHCP release [Draw-EM (92)], as transient zones in the fetal brain did not have to be distinguished for this step. The 5TT map is a 4D image where the last dimension comprises five mutually exclusive channels: cortical gray matter, subcortical gray matter, WM, cerebrospinal fluid (CSF), and pathological tissue (which did not exist for the normal fetal brains used in this work). We constructed the 5TT maps by preserving original cortical GM and WM labels, inserting an empty lesion channel, reassigning the brainstem and hippocampus to the subcortical GM channel, and including the lateral ventricles in the CSF channel. Because the original tissue segmentation sometimes overestimated cortical GM and extended into the subplate, it could create artificially closed gyral crowns and misplace the GMWMI. In these regions, streamlines could prematurely terminate within the gyrus or fail to reach the intended GMWMI. We therefore refined the tissue segmentation to restore anatomically appropriate tissue boundaries. This refinement expanded anatomically permissible tracking space where the interface had been incorrectly blocked, but did not impose streamline reconstruction; propagation remained determined by the local diffusion signal and tracking constraints. These corrections were critical to allow streamlines to traverse the expected anatomical zones such as the gyral crown and external capsule (figs. S1 and S2). In detail, we dilated the WM mask by one voxel (∼0.5 mm), identified its overlap with the cortical GM label, and reassigned the overlapping regions to WM. A similar correction was applied to the subcortical GM label, particularly to restore the intervening WM structures in the insular region. Given the variability in the size and shape of different superficial fiber bundles, a single tractography setting was not adequate to reconstruct all bundles. Therefore, we adopted an ensemble tractography strategy, using 10 distinct parameter configurations to compute the tractography. Each setting in the ensemble used a different combination of parameters, with angular thresholds selected between 70° and 90°, FOD amplitude cutoffs between 10−9 and 10−3, and FOD powers between 3 and 5. The ensemble design sampled a wide range of tractography settings to reduce dependence on any single parameter configuration and improve topology-level coverage and pathway continuity. All tractograms were generated using a probabilistic tracking algorithm (iFOD2) (27) with the sharpened DTI-derived dODF as the input. The minimum and maximum fiber lengths were set at 3 and 43 mm, respectively. For each of the 10 parameter configurations in the ensemble, 100,000 streamlines were generated. Therefore, the final ensemble tractogram obtained by merging the 10 ensemble results consisted of 1 million streamlines. To contextualize the local modeling and tractography choices used in the proposed pipeline, we provide representative results from the tractography configuration-selection procedure in fig. S2 and additionally performed a targeted qualitative comparison with two representative alternatives: deterministic tractography combined with ball-and-stick modeling and with generalized q-sampling imaging. The corresponding results are provided in fig. S3. To isolate the superficial fiber system, we applied a biologically informed exclusion strategy. In all exclusion steps below, streamlines were discarded in their entirety (i.e., not truncated at mask boundaries) if they intersected an exclusion mask (either endpoint within the mask, or any point along the streamline traversing it). We first excluded streamlines intersecting subcortical gray matter, brainstem, or cerebellum (to remove projection pathways). Deep WM fibers were then excluded using a developmental stage-specific strategy based on the presence of transient zones (to remove long association pathways). For fetuses younger than 32 weeks, deep WM was defined as the intermediate zone; only streamlines confined within the subplate were retained. The intermediate zone mask was eroded by one voxel (∼0.5 mm) to mitigate registration and partial volume errors. For fetuses older than 32 weeks, where the subplate is no longer separately labeled, deep WM was defined as the eroded WM core (two iterations, around 1 to 2 mm below the CP). This age-adaptive exclusion strategy ensured consistent delineation of the superficial system across developmental stages. The definition of the exclusion mask used to isolate the superficial tractogram from the whole brain streamlines is illustrated in Fig. 8. Subplate anatomy and exclusion mask for SAF tractogram isolation. Anatomical localization of the subplate (SP) and SAFs is shown by acetylcholinesterase (AChE) histochemistry (A and D) and in utero MRI (B, C, E, and F) of the fetal brain at very preterm [28 GW; (A) to (C)] and late preterm [36 to 37 GW; (D) to (F)] stages. (A) During the early preterm stage, the subplate is the largest compartment of the cerebral wall (96). The most superficial part of the subplate shows intense AChE reactivity and hyperintensity on T2-weighted MRI (asterisk). CP, cortical plate; WMS II, intermediate zone, consisting of sagittal strata and the periventricular crossroad of pathways; WMS I, deep periventricular zone. (D) At later stages, the subplate diminishes in size and its former position is progressively occupied by WMS III (centrum semiovale) and newly developing gyral white matter. Subplate dissolution is most prominent at the sulcal fundi (arrows). Dashed lines delineate the border between the subplate and underlying centrum semiovale. [(B) and (E)] In utero fiber extraction masks derived from T2-weighted MRI isolate superficial fibers by retaining only those confined to the subplate (or subplate remnant) and gyral white matter. Fibers traversing or terminating in excluded regions (red), including the centrum semiovale, subcortical gray matter, and brainstem, are removed. [(C) and (F)] Reconstructed SAFs extracted from the filtered superficial tractograms using the corresponding masks in (B) and (E). Gyri-level SAF bundle parcellation and reproducibility assessment Following the extraction of SAFs constrained to the superficial compartment, we applied a hierarchical evaluation procedure (algorithm S1). At the individual level, candidate bundles were extracted from the individual-specific tractograms using a parcellation-driven approach based on a gestational age–matched fetal cortical atlas (85). Streamlines were grouped by the gyri they connect (Fig. 1D), yielding approximately 300 intra- and intergyri candidate bundles per brain. A full list of the gyri used to define candidate connections is provided in table S1. Pairwise bundle connectivity at 36 GW is visualized in fig. S11. At the individual level, we investigated an outlier removal method that scores each streamline in a bundle by its similarity to the other streamlines in the bundle, i.e., cluster confidence index (93). Streamlines falling below the 3rd or 5th percentile of the bundle-specific score distribution were flagged as potentially spurious and removed. We validated this pruning strategy with the GA35 group to be consistent with the histological correspondences. Whereas visual inspection confirmed that pruning effectively eliminated irregular trajectories in individuals with low image quality, in most individuals, the initial unpruned bundles already showed coherent trajectories; pruning in these cases primarily removed peripheral streamlines at the edges of tube-like bundles. To preserve the full extent and anatomical shape of each connection, we ultimately retained the original, unpruned bundles for all individuals in downstream analyses. Bundles with fewer than 100 streamlines were excluded to ensure robustness in later group-level evaluation. To assess reproducibility across individuals, we performed nonlinear registration of each individual’s diffusion space to their age-matched template in an atlas reconstructed from the dHCP dataset (94). In this common space, we computed the binary mask for each bundle by thresholding its streamline density map at the 5th percentile of voxel-wise streamline density. Pairwise Dice coefficients were then calculated between each fetus’ bundle mask and the group-consensus bundle mask that was obtained by arithmetic averaging. To identify bundles with consistent anatomical presence across individuals, a two-step quality control procedure was applied. First, individual bundles were retained if their Dice coefficient of the overlap with the group-averaged bundle mask exceeded a minimum threshold (Dice ≥0.1). This liberal threshold was chosen solely to exclude invalid cases (e.g., empty or NaN masks) rather than to assess anatomical precision. Second, a bundle was considered group-wise valid if more than half of the individuals in the group exhibited a valid instance. This criterion ensured that retained bundles reflected consistent, nonrandom anatomical features rather than individual-specific artifacts. Following this initial filtering, we further refined the group composition by performing iterative individual-level outlier removal. This step aimed to exclude individuals whose bundles globally deviate from the rest of the group, thereby improving the overall group coherence. In each iteration, the individual whose exclusion most improved the mean Dice coefficient across bundles was removed. This process continued until improvements were no longer statistically significant (P > 0.01), negligible (improvement in average Dice coefficient ΔDiceˆ<0.005), or the group size fell below five individuals. Using the final retained set of individuals for each age group, we recomputed group-consensus bundle masks and identified the reproducible short association connections. In each age group, approximately 250 bundles passed quality control and were retained for bundle-level microstructural evaluation and developmental milestone analysis. Representative common-space results illustrating bundle reproducibility, cross-individual spatial consistency, and the corresponding group-consensus masks used for bundle-wise diffusion quantification are provided in fig. S4. Across the full cohort, 243 fetal MRI sessions from the publicly released dHCP dataset were initially considered. Sessions were included based on neuroradiological assessment of normative development and the availability of the essential imaging data required for SWM reconstruction, i.e., T2w MRI and diffusion MRI, resulting in 241 eligible sessions. The sample size for downstream regression analyses varied at the bundle level by the hierarchical quality-control workflow described above; after iterative outlier removal within age groups, 201 sessions were retained for regression, although the exact number of datapoints varied slightly across bundles and hemispheres. To facilitate targeted analysis and visualization, we selected a subset of SAF bundles that connect major gyri for detailed evaluation based on their developmental relevance and consistency across individuals. Specifically, we included the connection between the precentral and postcentral gyri, which are from the primary region of cerebral development and among the earliest-developing short association pathways. Additional ROIs were defined around early-forming sulci, such as the Sylvian fissure and calcarine sulcus, which are typically identifiable by 23 to 24 weeks of gestation. Also, these bundles have been reported for high reproducibility in prior SAF studies on adults (95). These regions were mapped onto the cortical parcellation atlas and used to extract and visualize representative bundles from our tractography results across gestational age groups. Individual reconstructions across multiple SAF systems, individuals, and developmental stages are provided in fig. S5 to complement the cohort-level occurrence analysis. Bundle-specific developmental modeling Free water–eliminated DTI was computed using a two-compartment model consisting of an isotropic compartment to represent the free water and an anisotropic compartment modeled with a standard diffusion tensor (91). FA and MD were computed from the diffusion tensor compartment. All measures were computed in each individual’s native space. Bundle volumes were estimated from native individual bundle masks defined by thresholding the voxel-wise streamline density map at the 5th percentile; in practice, this corresponded to retaining voxels traversed by at least one streamline. DTI measures, including FA and MD (or their free water–corrected counterparts), were summarized as the median value within group-consensus bundle masks warped back to each individual. Gestational age was treated as a continuous floating-point variable to preserve temporal resolution. Regression modeling of developmental trajectories was implemented as follows. For each bundle, linear and quadratic models were fit to the observed age-related data. Model selection was based on the adjusted coefficient of determination (Radj2), which penalizes model complexity. Subsequently, the selected model and nonparametric bootstrapping (1000 iterations) were used to estimate uncertainty bounds. In each iteration, a sample of the same size as the original dataset was taken from the dataset with replacement, and the selected model was fit to the bootstrap sample. This procedure resulted in 1000 regression coefficients and predictions. At each gestational age, the 95% confidence interval was computed as the 2.5th and 97.5th percentiles of the bootstrapped predictions. The resulting bounds characterize the uncertainty in the estimated mean trajectory, rather than the variability of individual observations. For bundles that were successfully reconstructed in both hemispheres, bilateral data were combined for regression but plotted using distinct colors to preserve lateralization information.