A healthy 22-week fetal brain has large fluid-filled ventricles and an almost smooth cortex. At 36 weeks, those same features indicate hydrocephalus or lissencephaly. The standard automated approach measures a structure and compares it with a growth curve at the fetus's recorded gestational age. That approach depends on the recorded age being right and the measurement being clean, and in multi-center fetal MRI neither is reliable. We propose a different test. Each anatomical measurement is converted into the age it implies, which we call its "structural age." The brain is then flagged as atypical when its structures disagree about how old it is, and the explanation names the structure that disagrees. A small young brain has consistent structural ages. A brain with enlarged ventricles but mature folding does not, whatever age is written on the requisition.
The idea joins four threads. The first is developmental anatomical variability: brain anatomy changes so fast that a fixed notion of "normal" does not hold. The second is morphological mimicry: normal immaturity looks like disease. The third is interpretability: the prediction passes through a small set of named, measurable quantities, a concept bottleneck, meaning a model forced to predict human-readable attributes before its final answer. The fourth is domain shift: scanner, site and image quality change the measurements themselves. We use development as the mechanism, mimicry as the target failure, a concept bottleneck as the form of the explanation, and multi-site fetal MRI as the stress test.
Core Claim
We claim that disagreement among structure-specific ages is a better signal of dysmorphology than deviation from recorded gestational age, for three reasons.
- Robustness to wrong age. It does not require the recorded age to be correct.
- Robustness to site. It partly cancels site-level measurement biases that shift all structures together.
- Interpretable output. It yields an explanation of the form "ventricles read five weeks younger than cortex, brain volume and cerebellum" without any concept-level annotation, because the concepts are measured from segmentations rather than labeled by experts.
Relation to Prior Work
The closest existing concept-bottleneck work is static. Concept Bottleneck Models introduced the image-to-concepts-to-label structure. A Textbook Remedy for Domain Shifts (KnoBo) grounds a vision transformer's concepts in textbook knowledge to resist site shortcuts on chest X-ray and dermoscopy. Coherent Concept-based Explanations spatially constrains dermoscopic concepts with a lesion mask. All three motivate our output format, but they assume a concept means the same thing for every patient. That assumption fails when "large ventricles" is normal at one age and pathological at another. They also require concept labels, or a text-derived concept vocabulary, that do not exist for fetal MRI.
On the developmental side, CAS-Net builds age-conditioned atlases, but it assumes neurotypical development and a known age. The work we build on most directly is Fetal gestational age prediction via shape descriptors of cortical development. It predicts age from cortical folding descriptors and observes that spina bifida brains systematically "read" as older or dysmorphic. That paper uses a single structure's age gap against chronological age. We generalize the gap to many structures and replace chronological age with their mutual agreement. The generalization matters most when chronological age is unreliable.
For the measurement layer we reuse the contrast- and pathology-agnostic SynthSeg fetal pipeline. It trains on images synthesized from label maps and improves pathological Dice from 0.60 to 0.67, at a small cost on neurotypical brains. Its limit is that it is a black box that segments without saying whether an enlarged ventricle is immaturity, artifact or disease. Two FeTA reports constrain our design:
- FeTA 2022 and FeTA 2024 show that Dice hides topological breakage.
- FeTA 2024 also shows that image quality and site move Dice more than pathology does.
We therefore treat segmentation topology as a quality signal that must gate the concepts. Finally, Improved post-hoc probability calibration for out-of-domain MRI segmentation uses a stationary cardiac shape prior and explicitly cannot handle changing geometry. Our age-indexed growth curves are a developmental replacement for that stationary prior.
Research Program
Measurement layer. A contrast-agnostic segmenter, the SynthSeg variant above, produces the seven FeTA tissue classes. From each segmentation we compute a handful of concepts that clinicians already reason with: lateral ventricle volume normalized by brain volume, supratentorial brain volume, cerebellar volume, extra-axial fluid volume, and a cortical folding descriptor. The folding descriptor is GSHOT where surfaces allow, with a simpler gyrification index as fallback.
Growth curves. On neurotypical scans only, we fit a normative curve for each concept $k$: the mean $\mu_k(a)$ and spread $s_k(a)$ as smooth functions of age $a$. The training scans are dHCP fetal, the CRL atlas, and FeTA neurotypical cases from the training sites.
Structural ages and the disagreement score. For a new scan with measured concepts $x_k$, the likelihood of age $a$ under concept $k$ gives an implied age $\hat a_k$ with uncertainty $\sigma_k$. These are computed from the likelihood rather than by inverting the curve, so that saturating curves, such as late folding, simply produce wide uncertainty. We then fit one shared latent age $a^\*$, with the recorded gestational age $g$ entering only as a weak prior of width $\sigma_g$:
$$a^\* = \arg\min_a \sum_k \frac{(\hat a_k - a)^2}{\sigma_k^2 q_k} + \frac{(g - a)^2}{\sigma_g^2}$$
Here $q_k \ge 1$ inflates the uncertainty of concept $k$ when its source segmentation is topologically implausible. We measure that implausibility by the per-structure Euler characteristic deviation used in FeTA 2024. The atypicality score is the weighted residual
$$D = \sum_k \frac{(\hat a_k - a^\*)^2}{\sigma_k^2 q_k}$$
The explanation for a scan is its list of per-concept residuals, each with a sign ("reads younger" or "reads older").
The $g$ term is kept deliberately. Some disorders shrink the whole brain coherently, as in global growth restriction. The structures then agree with one another but not with the calendar, and only the prior can catch that. The analysis therefore reports whether each detection came from inter-structure disagreement or from disagreement with recorded age.
What is built. The deliverable is the scoring pipeline, a released set of normative curves, and a benchmark of mimic cases described below. No new deep architecture is trained beyond the segmenter. This is intentional: the scientific question is whether the relative-age signal carries the information. If it does, the score can later be made differentiable and trained end to end.
Baselines
Naive reference: recorded-age z-score. This is the simplest possible check and the one closest to clinical practice. We fit the same growth curves and compute each concept's z-score at recorded gestational age, scoring the scan by the sum of squared z-scores. It uses identical segmentations and concepts and differs from our method only in replacing $a^\*$ with $g$. It is therefore the cleanest possible comparison, and it doubles as the key ablation. Our method beats it if it matches its pathology-detection area under the ROC curve on clean data and outperforms it when recorded age is perturbed or when the test site differs.
Single-concept age gap. Following the GSHOT shape-descriptor paper, we regress age from folding descriptors and flag scans with large residuals against recorded age. That paper reports R² of 0.89 and a mean absolute error of 6 days on neurotypical fetuses, against 10 days on spina bifida fetuses. This makes it the strongest existing age-based atypicality signal in the file. We reimplement it on the same surfaces, with SVR hyperparameters tuned by the same cross-validation budget as our curves.
Black-box image classifier. A 3D classifier, initialized from the nnU-Net encoder of the segmenter, is trained to separate neurotypical from pathological scans on the same training sites with matched augmentation and tuning budget. It represents what a practitioner would try first. No reported result was found on Lacuna for pathology classification on FeTA, so this baseline's number will come from our own runs.
Static concept bottleneck with age as input. A logistic model on the raw concepts plus gestational age is a KnoBo-style static bottleneck without text priors. It tests whether "concepts plus age" already suffices without the structural-age transformation.
Two further ablations. Setting all $q_k = 1$ isolates the topology gate. Dropping the $g$ term entirely isolates pure inter-structure agreement.
Why not just…?
Why not just compare each measurement against a growth curve at the recorded age? This is the real threat. On clean, correctly dated, single-site data it may perform as well as our method, and we expect roughly that. The difference should appear under three stresses:
- Dating error. Recorded age can be off by more than a week, and a correctly developing young fetus with an overstated age reads as uniformly "too small" to the z-score but not to the disagreement score.
- Site shift. Shifts that scale every volume at once also cancel in inter-structure comparison.
- Mimic cases. Enlarged ventricles at a later age are exactly the mimics of immaturity.
The age-perturbation and cross-site experiments below settle this directly. If our method shows no advantage there, the simpler z-score should be preferred and we will say so.
Why not just train a classifier on images with heavy augmentation? nnU-Net with augmentation won both FeTA challenges, yet FeTA 2024 still reports Dice drops of about 0.10 across image quality and 0.075 across sites. Its biometry models reached a mean absolute percentage error (MAPE) of 7.72%, against 9.56% for a gestational-age-only regression and 5.38% inter-rater variability (FeTA 2024). This suggests image models learn age and intensity shortcuts rather than structure. A pathology classifier trained on a few dozen pathological cases is likely to latch onto "small brain equals abnormal." Our matched-size mimic set tests exactly this.
Why not just fine-tune on a few scans from each new site? The cross-age, cross-site diffusion study shows fine-tuning on one or two target subjects beats harmonization, which is a strong result. However, it concerns signal estimation in normal development. Detecting pathology at a new site would require labeled pathological cases from that site, which are rare. Where a few annotated target scans exist, we will fine-tune the segmenter for all methods alike. Any remaining difference then comes from the scoring rule, not from adaptation.
Why not just use age-binned atlases? Binning makes the recorded-age dependence worse, not better, and introduces errors at bin edges (A Feature-Based Developmental Model of the Infant Brain). Our continuous curves subsume binning as a special case.
Proposed Experiments
Data. Growth curves are fit on dHCP fetal (274 scans), the CRL atlas, and neurotypical FeTA cases from Kispi and Vienna. Pathology detection is evaluated on FeTA pathological cases and on the 33-subject severe-malformation set from the SynthSeg study. We will use the neurotypical/pathological labels and gestational ages as they are distributed with FeTA; where they are unavailable, we restrict to the subsets with clinical metadata. Cross-site testing holds out CHUV, UCSF and the KCL 0.55T scans.
Metrics. Pathology detection is measured by area under the ROC curve and sensitivity at 90% specificity. Explanation correctness is the fraction of pathological cases whose largest-residual concept matches the diagnosed structure, for example the ventricles in ventriculomegaly. We also report segmentation Dice and Euler characteristic for context.
Age-perturbation test. We add ±1, 2 and 3 weeks of error to recorded gestational age. The prediction is that the z-score baseline's false-positive rate on neurotypical brains rises steeply, while the disagreement score's rises slowly.
Mimic test. We pair neurotypical fetuses at 22–25 weeks with ventriculomegaly cases at 30–36 weeks whose absolute ventricular volume matches. We supplement these with controlled synthetic ventricular dilation of late neurotypical label maps, as in the SynthSeg study. Success means the disagreement score separates the pairs, whereas the black-box classifier and raw-concept models do not beat chance by much.
Cross-site test. We compare the drop in detection AUC from training sites to held-out sites for every method. We separately check whether the topology gate reduces false positives on low-quality and 0.55T scans.
Coherent-shrinkage check. On cases with global growth restriction, if available, we verify that detection comes through the recorded-age term. We report honestly that pure inter-structure agreement cannot see such cases.
Failure criterion. If the disagreement score beats the recorded-age z-score neither under age perturbation nor cross-site, the core claim is refuted.
References
- Concept Bottleneck Models
- A Textbook Remedy for Domain Shifts: Knowledge Priors for Medical Image Analysis
- Coherent Concept-based Explanations in Medical Image and Its Application to Skin Lesion Diagnosis
- CAS-Net: Conditional Atlas Generation and Brain Segmentation for Fetal MRI
- Fetal gestational age prediction via shape descriptors of cortical development
- Towards contrast- and pathology-agnostic clinical fetal brain MRI segmentation using SynthSeg
- Multi-Center Fetal Brain Tissue Annotation (FeTA) Challenge 2022 Results
- Advances in automated fetal brain MRI segmentation and biometry: Insights from the FeTA 2024 challenge
- Improved Post-hoc Probability Calibration for Out-of-Domain MRI Segmentation
- Cross-Age and Cross-Site Domain Shift Impacts on Deep Learning-Based White Matter Fiber Estimation in Newborn and Baby Brains
- A Feature-Based Developmental Model of the Infant Brain in Structural MRI