Fetal brain wiring appears to accelerate sharply over a few weeks of gestation. The largest in-utero diffusion MRI study to date fit an S-shaped growth curve to 198 fetuses and placed a rapid strengthening of structural connectivity at roughly 27.5–30.5 gestational weeks (Charting the Normal Development of Structural Brain Connectivity in Utero using Diffusion MRI). Diffusion MRI tracks how water moves along nerve fibers. The standard fast readout, Cartesian echo-planar imaging (EPI), fills the image's frequency grid line by line. That design breaks badly when the subject moves, and it warps geometry when the magnetic field is uneven. Fetal motion and fetal anatomy both change with gestational age, so the size of these artifacts plausibly changes with age as well. An age-dependent artifact can bend a growth curve. The field currently has no way to tell whether the reported inflection point is biology, measurement drift, or both.
We propose a two-cohort design built to answer that question. A small cohort of about 30 fetuses is scanned twice in the same session: once with the standard Cartesian protocol and once with a motion-robust radial diffusion acquisition. In a radial acquisition every readout line (a "spoke") passes through the center of frequency space, so motion shows up as diffuse streaks rather than coherent ghosts, and the repeatedly sampled center can be used to track the motion. The paired scans give a direct, within-subject estimate of how Cartesian crossing-fiber measurements are biased as a function of motion and gestational age. That bias model is then transported to the large existing Cartesian cohort to re-estimate the growth curve. The small cohort is never asked to estimate the trajectory itself, a job that a sample of 30 is known to do badly. It is asked only to calibrate the measurement.
Core Claim
A small paired, motion-robust calibration substudy can correct the growth trajectory estimated from a large motion-sensitive cohort. It can also test whether the 27.5–30.5 week connectivity spurt survives once gestational-age-dependent measurement bias is removed. This is a measurement-error design, sometimes called regression calibration with an internal validation sample, applied to non-linear brain development.
The model can be stated compactly. Let $y_{i}$ be a crossing-fiber–derived connectivity measure for fetus $i$ at gestational age $a_i$, and let $m_i$ be a vector of motion and distortion summaries for that scan. We model the Cartesian measurement as
$$y_i^{\text{cart}} = f(a_i;\theta) + b(m_i, a_i;\phi) + \varepsilon_i ,$$
and the radial measurement as $y_i^{\text{rad}} = f(a_i;\theta) + \eta_i$. Here $f$ is a sigmoid growth curve whose parameters $\theta$ include the inflection age $a^\ast$. The term $b$ is a low-dimensional bias function with parameters $\phi$, and $\varepsilon_i, \eta_i$ are noise terms. The paired cohort identifies $\phi$ from within-subject differences $y^{\text{cart}}-y^{\text{rad}}$. These differences cancel $f$ exactly, which is why they need far fewer subjects than the curve itself. The large cohort then estimates $\theta$ with $b$ subtracted, and uncertainty in $\phi$ is propagated by bootstrap.
The approach has limits. The claim rests on one assumption: the radial measurement is approximately unbiased with respect to motion. We test this assumption in simulation before relying on it. The design also has a failure mode we would report. If a simple motion-covariate model on Cartesian data gives the same corrected curve, the radial arm was unnecessary.
Relation to Prior Work
Several lines of prior work motivate this study and supply its methods.
The in-utero connectome study above is the target we re-examine. It uses single-shell spherical deconvolution, ensemble tractography with SIFT2 weighting, and a bootstrapped sigmoid fit on Cartesian EPI after slice-to-volume reconstruction. It does not model how residual artifacts vary with gestational age, and that is the gap we fill.
The deconvolution method comes from Ground-Truth Effects in Learning-Based Fiber Orientation Distribution Estimation in Neonatal Brains. Spherical deconvolution estimates, in each voxel, a fiber orientation distribution: a function on the sphere whose peaks are fiber directions. That paper shows that in immature white matter, single-shell three-tissue deconvolution (SS3T-CSD) finds multiple fiber directions in about 61% of voxels, while the adult-standard multi-shell multi-tissue variant finds them in only 23%. It also shows that its learned estimator holds up across the 35-to-42-week developmental shift. We adopt SS3T-CSD as the default estimator. Because it works from a single shell of measurements, it also fits the short radial protocol.
A second estimator is Equivariant spatio-hemispherical networks for diffusion MRI deconvolution. It is a rotation-equivariant network that exploits the symmetry of diffusion signals and handles sparse gradient-direction sampling. We use it because the radial protocol will have few directions. That paper worked only on Cartesian adult and synthetic data.
The radial side builds on two papers. Rapid, high‐resolution and distortion‐free R2∗ mapping of fetal brain using multi‐echo radial FLASH and model‐based reconstruction showed that radial imaging of the moving fetal brain is feasible and distortion-free. It reported a scan-rescan intraclass correlation of 0.9213 under motion in 9 fetuses, but it measured relaxation rather than diffusion. Learning Apparent Diffusion Coefficient Maps from Undersampled Radial k-Space Diffusion-Weighted MRI in Mice using a Deep CNN-Transformer Model in Conjunction with a Monoexponential Model showed that radial diffusion data undersampled 8x (50 of 403 views) can still yield scalar diffusion maps, with SSIM 0.9937 versus 0.8486 for least-squares fitting. It measured only scalar diffusion, not fiber orientations, and reported fully sampled radial diffusion scans of up to 25 minutes per slice. That figure is the main constraint on our design: the radial arm must be short, low in direction count, and heavily undersampled.
Two papers constrain the statistical design. Reproducible brain-wide association studies require thousands of individuals shows that at N≈25, significant effects inflate by more than 200% and signs flip between random draws. This is why the small cohort must not be used to estimate the trajectory. Mapping the basal ganglia alterations in children chronically exposed to manganese reports that N=23 made non-linear age modeling infeasible. We take this as direct evidence that small developmental cohorts need a different role.
Predicting Infant Brain Connectivity with Federated Multi-Trajectory GNNs using Scarce Data attacks data scarcity with imputation and pooling across sites. That approach reduces variance, whereas ours targets bias. The two are complementary.
Research Program
The work proceeds in three stages. Each stage can stop the project if its result is negative.
Stage 1: simulation. We start from dHCP neonatal diffusion data (The Developing Human Connectome Project Neonatal Data Release, 758 sessions with diffusion MRI passing QC). We impose a known sigmoid on synthetic connectivity by re-weighting, then pass images through two forward models. One is a Cartesian EPI model with motion-induced phase errors, signal dropout, and susceptibility warping. The other is a golden-angle radial multi-shot model with the same motion traces, reconstructed with center-of-k-space motion and phase navigation plus learned undersampled reconstruction. Golden-angle ordering rotates each spoke by about 111°, so any window of spokes covers angles evenly and corrupted spokes can be discarded without leaving gaps. Motion amplitude is made to decline with gestational age, a stated assumption that we vary in sensitivity analyses. We also inject orientation ground truth from the DiSCo phantom so angular errors are known. This stage asks three questions. Does an age-dependent motion profile shift the fitted inflection age? By how much? Is the radial pathway's bias small enough to serve as a reference?
Stage 2: acquisition and paired calibration. We build a single-shell, roughly 30-direction golden-angle radial diffusion protocol. Each spoke's central sample serves as a navigator for correcting shot-to-shot phase. The protocol is time-budgeted to fit alongside a dHCP-matched Cartesian protocol in one session. We first pilot it in about 10 neonates, where motion is milder, and then scan about 30 fetuses spread evenly across 22–37 weeks. Both arms are processed identically downstream. Paired within-subject differences in fiber orientation distributions, crossing-fiber fraction, and tract-level connectivity are regressed on motion summaries and gestational age to estimate the bias function $b$.
Stage 3: re-estimation. We apply the bias model to the 198-fetus Cartesian cohort and re-fit the sigmoid. Error is propagated jointly from both cohorts by a nested bootstrap. We report the corrected inflection age and spurt width with intervals.
Baselines
The naive reference point reproduces the published analysis: SS3T-style single-shell deconvolution, ensemble tractography, SIFT2 weighting, and a sigmoid fit on the Cartesian fetal cohort, with no bias correction. The reported result is a spurt at approximately 27.5–30.5 gestational weeks (Charting the Normal Development of Structural Brain Connectivity in Utero using Diffusion MRI). Our corrected estimate is compared against this window using the same subjects, parcellation, and tractography parameters.
The strongest competitor is covariate adjustment. It adds per-scan motion and distortion summaries, drawn from slice-to-volume reconstruction outputs and outlier-slice counts, as nuisance terms in the same sigmoid model. It is also given a smooth age-by-motion interaction and fitted as a hierarchical model with shrinkage. This is the cheap alternative a reviewer would reach for, so it gets the same bootstrap, the same tuning budget, and access to every motion summary our bias model uses. We count as beating it only if, in simulation where the true inflection is known, our corrected inflection age has lower bias and correct interval coverage while covariate adjustment does not.
For deconvolution quality we compare SS3T-CSD, multi-shell multi-tissue CSD, and the equivariant hemispherical network on both arms. The reference numbers available are 61% versus 23% multi-fiber voxels, and 70–76% multi-fiber agreement for an SS3T-trained U-Net with 28–45 input directions (Ground-Truth Effects in Learning-Based Fiber Orientation Distribution Estimation in Neonatal Brains). For the equivariant network, no reported angular-error or Tractometer result was found on Lacuna. We will report its angular error on DiSCo ourselves.
Two ablations isolate the key ingredients. The first replaces the radial reference with a second Cartesian scan in the same session. This yields test-retest calibration, which captures random but not systematic error, and so isolates whether motion-robust sampling is doing the work. The second fits the sigmoid on the 30-fetus radial cohort alone. We expect this to give wide, unstable inflection estimates, which shows why the small cohort is used only for calibration.
Why not just…?
Why not just add motion covariates to the Cartesian growth model? This is a real threat, and Stage 1 is designed to settle it. Covariate adjustment corrects only for the artifact that the recorded motion summaries capture. It also cannot separate motion from age when the two are strongly collinear, which is the situation we suspect. Susceptibility warping and dropout in crossing regions may leave biased orientation distributions even in scans with modest recorded motion. If the simulation shows that covariate adjustment recovers the true inflection as well as paired calibration does, we will report that and drop the radial arm from future recommendations.
Why not just exclude high-motion scans? If motion declines with gestational age, exclusion removes young fetuses preferentially. That thins the already sparse 10–15 subjects per week at the early end of the curve and biases the fitted inflection toward later ages. We will run exclusion at several thresholds as an additional comparison and report the inflection shift together with the number of scans retained.
Why not just use a larger pooled Cartesian sample or federated normative models? More subjects reduce variance, not bias. The warning from Reproducible brain-wide association studies require thousands of individuals concerns sampling variability, but a systematic age-linked artifact persists at any N. Pooling in the style of Predicting Infant Brain Connectivity with Federated Multi-Trajectory GNNs using Scarce Data is compatible with our correction and could be layered on after it.
Why not just use tensor imaging with fewer directions to shorten the radial scan? A tensor fit averages crossing fibers into a flattened or spherical shape (Connectivity-Based Parcellation of the Cortical Mantle Using q-Ball Diffusion Imaging). The connectivity measure we are calibrating depends on crossing fibers. A tensor-based reference would therefore calibrate the wrong quantity.
Proposed Experiments and Decisive Checks
Check 1 (simulation). With a known sigmoid and an age-declining motion profile, we measure the Cartesian pipeline's inflection-age bias, the radial pipeline's bias, DiSCo angular error, and Tractometer valid-streamline fraction. The idea is worth pursuing only if two things hold: the Cartesian bias in inflection age is comparable to the width of the reported 3-week window, and the radial bias is substantially smaller. If Cartesian bias is negligible, the published spurt is vindicated and the project ends there with a useful negative result.
Check 2 (neonatal pilot). We verify that radial single-shell data give multi-fiber voxel fractions and paired orientation agreement with Cartesian data in the range the neonatal SS3T literature reports. We also verify that the time budget holds.
Check 3 (fetal paired cohort). We estimate $b$ and test whether it depends on gestational age beyond what the motion summaries explain. A null result here would favor covariate adjustment.
Check 4 (re-estimation). We re-fit the 198-fetus curve with the bias correction and compare the corrected inflection interval against 27.5–30.5 weeks. The covariate-adjusted, exclusion, test-retest-calibration, and radial-only fits are reported beside it.
The outcome is informative in either direction. Either the spurt survives calibration, or it moves or widens. In both cases the field gains a reusable template for letting a small, expensive, motion-robust cohort correct a large, cheap one.
References
- Charting the Normal Development of Structural Brain Connectivity in Utero using Diffusion MRI
- Ground-Truth Effects in Learning-Based Fiber Orientation Distribution Estimation in Neonatal Brains
- Equivariant spatio-hemispherical networks for diffusion MRI deconvolution
- Rapid, high‐resolution and distortion‐free R2∗ mapping of fetal brain using multi‐echo radial FLASH and model‐based reconstruction
- Learning Apparent Diffusion Coefficient Maps from Undersampled Radial k-Space Diffusion-Weighted MRI in Mice using a Deep CNN-Transformer Model in Conjunction with a Monoexponential Model
- Reproducible brain-wide association studies require thousands of individuals
- Mapping the basal ganglia alterations in children chronically exposed to manganese
- Predicting infant brain connectivity with federated multi-trajectory GNNs using scarce data
- The Developing Human Connectome Project Neonatal Data Release
- Connectivity-Based Parcellation of the Cortical Mantle Using q-Ball Diffusion Imaging