Age-Aware Left–Right Comparison for Segmenting the Developing Brain

Medical image models often find subtle damage by comparing the two sides of the body. The healthy side serves as a control matched to the same patient: if the left hemisphere looks different from the mirrored right hemisphere, something may be wrong. This works in adult brains because the two hemispheres are nearly mirror images and the small differences between them stay stable. In fetuses and newborns that assumption breaks. The brain is folding, myelinating and changing tissue contrast week by week, and the normal difference between the left and right sides changes along with it. A naive mirror comparison will therefore flag healthy developmental asymmetry as damage at some ages and miss real damage at others.

We propose that a symmetry-based segmentation model for the developing brain should compare each side against its mirrored partner and then measure that difference against the left–right difference expected at the scan's gestational age. The model has two output channels. The first is an asymmetry deviation: how far the observed left–right difference sits from what is normal at that age. The second is a bilateral deviation: how far both sides together sit from normal volumes for that age. The second channel exists because injury caused by body-wide factors, such as unstable blood flow, tends to affect both hemispheres. A pure left–right comparison cannot see that kind of injury.

Core Claim

Bilateral comparison helps volumetric segmentation and anomaly detection in developing anatomy only if normal asymmetry is modeled as a function of age. Without that, mirror-based attention adds age-dependent false positives that can cancel its benefit. With it, the healthy side becomes a usable control even while the brain is changing quickly. The result is lower subject-to-subject variance for one-sided findings than population growth charts give, and the bilateral channel still catches symmetric injury.

How the Directions Fit, and What Already Exists

The symmetry mechanism comes from adult work. SEAN aligns head CT scans to the midline with an unsupervised network. It then uses cross-hemispheric attention, in which each voxel looks at its mirrored counterpart, to segment early infarcts. It reached a Dice overlap of 0.5784 against 0.4588 for a U-Net on the Acute Ischemic Stroke Dataset (AISD). Contralaterally Enhanced Networks mirror candidate regions across the spine in chest X-rays and subtract shared anatomy. BrainNormalizer borrows edge maps from the contralateral hemisphere to inpaint pseudo-healthy brains. All three assume mature, fixed geometry. Our proposal reuses their mirror-attention idea and drops that assumption.

The developmental side supplies age conditioning. CAS-Net warps a template into an age-specific atlas and registers it to each fetal scan. It reached 85.2% mean Dice on dHCP fetal MRI against 70.1% for a 3D U-Net. UniCoN injects embryonic age and spatial coordinates into network bottlenecks. Neither model compares hemispheres, so we build on their conditioning and add the comparison they lack.

Several papers motivate the problem. Infant gray–white contrast inverts over the first two years (de Macedo Rodrigues et al.). Folding changes rapidly in preterm brains (prefrontal folding study). Adult templates misregister infant anatomy (Braga et al.).

Other papers constrain the design. Natural asymmetry produces false positives under strict mirror assumptions (Fineide et al.). Systemic disease that affects both sides defeats contralateral reference (Chen and Liu).

Blood pressure variability (BPV) enters as that systemic case, not as a model input. In adults, day-to-day and within-day swings in blood pressure predict silent infarcts independently of mean pressure (Chi et al., daytime systolic SD odds ratio 1.374). They also predict diffuse cognitive and vascular outcomes (Cui et al.). These are diffuse, often bilateral injuries, which is exactly what a left–right comparison misses and why the bilateral channel is required.

We considered and rejected the more obvious combination, feeding blood pressure time series into a segmentation network. No dataset in our material pairs blood pressure recordings with voxel labels. Volumetric segmentation is also defined by image content, and blood pressure has no causal path to a voxel's label. The physiology belongs in downstream validation, not in the segmenter.

Research Program

The model is a 3D segmentation backbone in the nnU-Net / CAS-Net family with three additions.

First, a midline alignment step based on SEAN's symmetry and restoration losses, retrained on fetal and neonatal MRI. Errors in midline estimation are reported as a separate failure mode.

Second, a mirror cross-attention block at the bottleneck. It lets features at each location attend to features at the reflected location.

Third, age conditioning of that block. Gestational or postmenstrual age is injected UniCoN-style, and the attention is computed on the residual between the observed left–right feature difference and a learned age-dependent expected difference.

To make the expected asymmetry concrete, let $d(x)$ be the feature difference between location $x$ and its mirror, and let $\mu_a(x)$ and $\sigma_a(x)$ be the mean and spread of $d(x)$ among healthy subjects of age $a$. The comparison signal passed to the decoder is

$$ z_a(x) = \frac{d(x) - \mu_a(x)}{\sigma_a(x)}. $$

$\mu_a$ and $\sigma_a$ come from a small network trained only on healthy-term and low-risk scans. The bilateral channel is the analogous age-normalized deviation of the summed left and right features.

We will first run an analysis that requires no new model. Using the existing dHCP automated tissue segmentations, split at the estimated midline, we will measure per-structure left–right volume asymmetry as a function of age. The question is whether that asymmetry is large and age-varying relative to scan-to-scan noise. If it is flat and small, the age-aware term is unnecessary and the project reduces to applying SEAN to infants. That would be a useful negative result, and it would be cheap to obtain.

Baselines

The naive reference is a plain 3D U-Net trained on the same dHCP fetal split. CAS-Net reports 70.1% mean Dice for this model on that data, and we will reproduce it to confirm our split matches.

The strongest existing segmenter is CAS-Net itself, at 85.2% mean Dice across nine tissue classes on dHCP fetal MRI. It is the right competitor because it already handles age-driven shape change through atlas conditioning. Any gain we show must come from the comparison between hemispheres, not from age conditioning alone. We will run it from the same preprocessing, the same subject-level split and the same GPU-hour tuning budget.

For the pathological fetal domain, the reference is the SynthSeg-based pipeline of Shang et al., which reports 0.67 Dice on severely abnormal cases against 0.60 for its baseline.

The direct symmetry competitor is SEAN-style mirror attention without age conditioning, retrained on infant data. Its only published result is 0.5784 Dice on adult AISD CT. We found no reported result on Lacuna for any symmetry-attention model on fetal or neonatal MRI.

Two ablations isolate the key ingredient. One keeps the mirror attention but sets $\mu_a=0$ and $\sigma_a=1$, giving a mirror model that ignores age. The other removes the bilateral channel.

Beating the field means three things. Our model must exceed CAS-Net's Dice on dHCP fetal by more than the across-seed standard deviation. It must match or exceed 0.67 on the pathological domain. And on detection of small one-sided lesions, it must clearly beat the mirror model that ignores age while keeping false positives on healthy scans flat across age bins.

Why not just…?

Why not just add a left–right flipped copy of the scan as an extra input channel to nnU-Net with age as a covariate? This is the most serious threat, because a large network might learn age-specific asymmetry on its own. The experiments settle it directly: we include this flipped-channel model as a baseline under the same budget. If it matches our detection curves and false-positive profile across age bins, the explicit normative asymmetry term is not needed and we will report that. We expect it to fall short mainly in the low-data regime, since sparse annotations are the norm in developmental imaging (UniCoN has under 3% labeled slices). We will therefore test it at 10%, 30% and 100% of the training labels.

Why not just use population growth charts of structure volumes, with no symmetry at all? Growth-chart normative modeling handles symmetric injury naturally, and that is why our bilateral channel exists. Its weakness is that between-subject variance in total volume is large, so small one-sided changes fall within normal limits. The paired comparison removes most of that variance, which is the core premise of bilateral analysis (Liu et al.). We test this premise as the smallest synthetic one-sided lesion volume each method detects at a fixed false-positive rate. If growth charts match the paired approach, the asymmetry branch has no value.

Why not just use CAS-Net and inspect left–right volume differences after segmentation? Asymmetry measured after segmentation inherits boundary errors on each side. In low-contrast neonatal cortex those errors can be as large as the asymmetry itself. Building the comparison into the features lets the network use intensity evidence that a label map discards. We run the post-hoc version as a baseline on the same lesion-detection task.

Proposed Experiments

Segmentation is evaluated on dHCP fetal MRI (274 scans, 20.6–38.2 weeks) and dHCP neonatal MRI (886 sessions with T2 scans that passed quality control) from the dHCP release, and on the FeTA multi-domain set used by Shang et al. That set spans in-domain, out-of-domain, 0.55T and pathological scans. Metrics are Dice and 95th-percentile Hausdorff distance per tissue class, reported per gestational-age bin so that age-dependent failures are visible rather than averaged away.

Anomaly detection uses synthetic lesions inserted into healthy dHCP scans. One-sided lesions are hypointense or hyperintense regions of controlled volume in white matter and cortex. Bilateral lesions are symmetric white-matter signal changes and ventricular dilation of the kind Shang et al. synthesize. Real pathology comes from the 33 abnormal FeTA subjects. Metrics are AUROC as a function of lesion volume, together with the false-positive rate on healthy scans by age bin.

The decisive check is a crossover pattern. In the age bins where the first-stage analysis finds large normal asymmetry, the mirror model that ignores age should produce more healthy-scan false positives than no-symmetry CAS-Net. The age-aware model should remove that excess. If the excess never appears, the developmental argument fails.

A second check covers coverage of symmetric injury. Bilateral synthetic lesions should be detected by the bilateral channel and largely missed by both mirror-only variants. This would confirm that the two channels behave differently.

The physiological validation is exploratory and stated as such. If the dHCP collateral clinical records for the preterm subset include perinatal hemodynamic measures, we will test whether the bilateral-deviation channel associates more strongly with those measures than the asymmetry channel does. That prediction follows from systemic hemodynamic injury being diffuse, as the adult blood pressure literature suggests. If such records are absent, this test requires a collaborating neonatal cohort, and no claim about blood pressure will be made from this work.