Metallurgists judge whether a deformed metal has been properly "reset" by looking at how uniform its internal structure is, not at its average grain size. A sheet can have a good mean grain size and still crack, because a strained surface layer sits over an unstrained core. We propose borrowing that habit for developing brains. Brain white matter is the wiring made of long fiber bundles, or tracts. As it matures, the tissue along each bundle becomes more organized, and we expect that organization to become more even from one end of the bundle to the other. Early-life adversity is known to shift the pace of maturation in opposite directions in boys and girls. Our hypothesis is that this shift appears as a difference in how uniform the microstructure is along specific commissural tracts, meaning the bundles that cross between hemispheres, and that it does not show up in their averages. Diffusion MRI, which measures how freely water moves in tissue and so how aligned and dense the fibers are, can measure that uniformity in the living brain.
The opportunity comes from three results that do not yet meet. The first is Dimensional adversity, brain-age, & mental health. It found that deprivation and low socioeconomic status go with younger-appearing brains in male adolescents and older-appearing brains in females in ABCD. Its measure was a single whole-brain age estimate from structural T1-weighted MRI, and it found no mediation of mental-health outcomes in the combined sample. The second is Deep learning with diffusion MRI as in vivo microscope. It showed that adult sex differences in microstructure concentrate in the corpus callosum and middle cerebellar peduncle, but only in 22–37 year olds. The third is Local Brain-Age: A U-Net Model, which showed that global age scores dilute regional effects, but only in adults and only on T1-derived volume maps. We ask whether the sex-divergent pacing found in the first study is carried by the tracts the second study identified, and we measure it with a distribution-level statistic rather than a mean.
Core Claim
For each major tract, we sample fractional anisotropy (FA, a 0–1 index of how directionally aligned water diffusion is) at $K$ evenly spaced nodes along the bundle. This gives a profile $f_i(s_k)$ for participant $i$. Each tract has a characteristic shape because of crossing fibers and curvature. To remove that shape, we subtract an age- and sex-specific normative profile $\mu_{a,g}(s_k)$ estimated from low-adversity participants. We then define a heterogeneity index
$$H_i = \frac{1}{\bar f_i}\sqrt{\frac{1}{K}\sum_{k=1}^{K}\big(f_i(s_k)-\mu_{a,g}(s_k)-\bar r_i\big)^2},$$
where $\bar f_i$ is the participant's mean FA along the tract and $\bar r_i$ is the mean of the residual $f_i-\mu_{a,g}$. $H_i$ is therefore a scale-free measure of residual non-uniformity, separate from the mean shift $\bar r_i$. For tracts cut by the midsagittal plane, such as the corpus callosum, we add a core-versus-edge contrast: the FA difference between the central and peripheral voxels of the tract's cross-section. This is the analogue of the surface-to-core strain gradient that cross-roll rolling is designed to remove (Effect of Cross-Roll Angle).
The claim has three parts. First, in normative data $H$ falls with age. Second, adversity dimensions move $H$ in opposite directions by sex in commissural tracts: persistently high $H$ in deprived males, precociously low $H$ in deprived females. Third, a tract-level microstructural age gap built from $H$ and $\bar r$ mediates the threat-to-externalizing link in females, which the global T1 age gap failed to do in the combined sample. What is new is the observable (uniformity rather than mean), the tissue (diffusion microstructure rather than T1 volume), and the age window (9–14 rather than adulthood).
Why the Materials Lens, Honestly
The metallurgy literature does not supply a biological mechanism, and we do not claim that axons recrystallize. It supplies a measurement habit and a hypothesis about sequence. Dynamic recrystallization studies describe a staged progression: dislocations tangle, then form subgrain walls, then new uniform grains appear. Success is judged by uniformity of the resulting population, for example the refined 0.6 µm grains reported for asymmetric gradient extrusion. The same literature shows that where reorganization starts depends on local sites such as particular grain boundaries (Dislocation Transformations at 30°⟨0001⟩ Grain Boundaries). That implies transient non-uniformity before homogenization completes.
The transferable prediction is that maturation passes through a phase of spatial non-uniformity. On this reading, "accelerated" and "delayed" maturation differ in when uniformity arrives, and a mean-based index cannot tell those apart. The lens earns its keep only if the first experiment below shows age-related homogenization in normative data. If it does not, we drop it.
Research Program
The primary cohort is ABCD. The adversity dimensions (threat, deprivation, household instability, socioeconomic status), the two-year imaging follow-up and the three-year Child Behavior Checklist outcomes are exactly those used by the dimensional adversity study, so its findings can be replicated within our pipeline. The cited paper used only T1-weighted scans. Using ABCD's diffusion acquisition is our addition and must be confirmed against the data release. We will reconstruct FA, mean diffusivity and, where the acquisition's multiple diffusion strengths allow, mean kurtosis, which measures departures from simple Gaussian diffusion and reflects tissue complexity. We will track the corpus callosum segments, the middle cerebellar peduncle and a set of association tracts that serve as non-commissural controls. We will then compute $\bar r$, $H$ and the core-edge contrast for each tract.
Site effects are handled in a way informed by Multi-cohort, multi-sequence harmonisation. That study found that aggressive latent-factor harmonization (RELIEF) weakened a true age association from $b=-0.37$ to $b=-0.28$, while parameter-aware ComBat preserved it. We therefore apply ComBat-style correction to the tract features with age, sex and adversity explicitly protected. We report every result both with and without this correction.
Heterogeneity statistics are especially vulnerable to confounds, so we control for head motion during the diffusion scan, intracranial volume and tract length. Motion differs by sex and by adversity, and brain size differs by sex; either could fake a sex-by-adversity effect on $H$.
The analyses run in three stages. First, we fit normative age curves of $H$ and $\bar r$ by sex in low-adversity youth. Second, we test sex × adversity interactions on each tract's features. Third, we train a small tract-level microstructural age model: penalized regression on $\{H,\bar r\}$ across tracts, fit on low-adversity participants. Its gap is then entered into the same moderated-mediation models the dimensional adversity study used. As an out-of-cohort anchor, we apply the index to HCP diffusion data. This checks whether, by adulthood, the commissural tracts that carry adolescent adversity effects are the same ones that dominate adult sex differences.
Baselines
The naive reference point is the one most likely to embarrass us: tract-mean FA from atlas-defined bundles, entered into linear models with sex × adversity terms. This is the standard region-of-interest analysis, run on the same participants, tracts, covariates and harmonization. The deep-learning diffusion study argues that such averages miss non-linear, localized patterns, but it never tested region-of-interest means on adolescent adversity. Our method beats this baseline only if $H$ yields interaction effects or mediation that survive correction for multiple comparisons where $\bar r$ does not. It must also add at least a pre-registered increment in explained variance of three-year externalizing scores in held-out ABCD sites.
The strongest existing method on this exact question is the global T1 age gap from a pre-trained simple fully convolutional brain-age network (SFCN-reg), as used in Dimensional adversity, brain-age, & mental health. We rerun it on the same subsample. That paper reports sex moderation qualitatively and reports that the age gap did not mediate adversity-to-outcome links in the combined sample. We found no numeric effect size for its moderation on Lacuna, so our replication supplies the reference numbers. Beating it means showing significant female-specific mediation of the threat-to-externalizing link through our tract-level gap under identical models.
Two stronger spatial competitors are run under matched tuning budgets and site-held-out splits. The first is a T1 local brain-age U-Net in the style of Local Brain-Age, retrained on pediatric data. That paper reports a median MAE (mean absolute error) of 9.5 years in adults, with no pediatric result found on Lacuna. The second is a 2D slice-wise CNN with an LSTM (a recurrent network that reads the slices in sequence) on FA maps. This follows Brain Age Estimation Using LSTM on Children's Brain MRI, which reports MAE 0.96 years internally and 1.14 years on an unseen hospital dataset in NIH-PD. For brain-age accuracy generally, Global-Local Transformer reports MAE 2.70 years over ages 0–97. That number is not comparable to the narrow 9–14 ABCD window, so we compare all models against a predict-the-mean baseline there.
The key ablation removes $H$ and the core-edge contrast from our age model and keeps only $\bar r$. It isolates whether uniformity, rather than tract choice or diffusion data, drives any gain.
Why not just…?
Why not just use atlas tract-mean FA in a regression? This is the real threat, and the ablation above settles it directly. If sex × adversity effects are equally strong in $\bar r$, the homogenization framing adds nothing, and we will report that the finding is simply "diffusion means in commissural tracts." The case for $H$ is the specific prediction that accelerated and delayed maturation can share a mean but differ in uniformity. That is falsifiable, since it requires participants with matched $\bar r$ but divergent $H$ by sex and adversity.
Why not just use the full along-tract profile in a penalized model? A full profile with dozens of nodes per tract across many tracts puts us back in the regime where features far outnumber participants. Multimodal Sparse Classifier showed accuracy dropping from 85.45% to 82.83% when more feature sets were concatenated. We will nonetheless run the full-profile elastic net as a baseline. If it predicts outcomes better, we will say so, and we will check whether its weights concentrate on high-heterogeneity segments. If they do, that supports $H$ as a compressed, interpretable summary rather than a competitor.
Why not just train a deep network on whole FA volumes, as the in vivo microscope work did? That work reached a test AUC of 0.98 for adult sex classification from FA maps. Its 3D ViT without pretraining fell below 0.80, and 3D CNNs generalized poorly across pediatric scanners (2.64 vs 1.14 years MAE in the LSTM study). More importantly, a classifier's saliency map does not answer the mediation question. We include the 2D+LSTM FA model as a baseline and test whether its age gap mediates outcomes. If it matches our index, the conclusion is that diffusion data matter more than the uniformity statistic.
Decisive Checks
The first check is normative. If $H$ does not decline with age across 9–14 in low-adversity youth in at least the corpus callosum, the homogenization hypothesis fails, and the program reduces to a diffusion replication of the dimensional adversity study.
The second check is specificity. Sex-opposite adversity effects on $H$ should be larger in commissural tracts than in association-tract controls, and they should survive motion and brain-size control. A uniform effect across all tracts would point to an acquisition artifact.
The third check is mediation. In site-held-out models, the tract-level gap should mediate threat-to-externalizing in females and not in males. Its incremental variance explained should exceed both the global T1 age gap and the tract-mean-only ablation.
The fourth check is harmonization robustness. Conclusions must hold with protected ComBat and without harmonization, and must weaken under RELIEF-style correction in the direction the cerebrovascular study predicts. That pattern would show the signal is biological rather than site-driven.
References
- Dimensional adversity, brain-age, & mental health: Differences in male and female adolescents
- Deep learning with diffusion MRI as in vivo microscope reveals sex-related differences in human white matter microstructure
- Local Brain-Age: A U-Net Model
- Effect of Cross-Roll Angle on Microstructures and Mechanical Properties during Cross-Roll Rolling in AZ31 Alloys
- Deformation Characteristics of Asymmetric Gradient Extrusion in Preparing Ultra-Fine-Grained Bulk Materials
- Dislocation Transformations at the Common 30°⟨0001⟩ Grain Boundaries During Plastic Deformation in Magnesium
- Multi-cohort, multi-sequence harmonisation for cerebrovascular brain age
- Brain Age Estimation Using LSTM on Children's Brain MRI
- Global-Local Transformer for Brain Age Estimation
- Multimodal Sparse Classifier for Adolescent Brain Age Prediction