A preterm infant scanned at term-equivalent age, the age at which they would have been born had the pregnancy gone to term, often has a brain that looks "younger" than a term-born baby's. The deep grey matter is smaller, the white matter is less myelinated, and the cortex is less differentiated. Clinicians and researchers would like a single, individual, spatially resolved answer to the question "how far behind is this brain, and where?" A natural way to get it is to train a neural network to read maturity from MRI, then search by gradient descent for the smallest anatomically plausible edit that makes the network read the scan as fully mature. The size of that edit is the delay, and where the edit happens is the map. Adversarial machine learning has shown, many times over, that this kind of search gives misleading answers when the network's output is saturated. A confident softmax produces nearly zero gradient, so the optimizer stalls and reports a small distance that reflects the optimizer, not the input.
This note proposes to import that lesson directly. We define a constrained "maturational distance" for individual preterm neonates. We show whether cross-entropy-based searches underestimate and mislocalize delay, and we test whether margin-based objectives, of the kind that fixed the same failure in robustness evaluation, recover delay patterns that agree with independent biological measurements. The deliverable is a measurement method with an explicit reliability audit. It is not a new brain-age predictor.
Core Claim
Gradient-based individual delay estimates from neonatal maturity models are only trustworthy if the search objective does not saturate. Our first claim is that a counterfactual search driven by a cross-entropy readout of binned postmenstrual age (PMA, the age counted from the mother's last menstrual period) will systematically underestimate delay. It will also concentrate its edits on coarse features such as ventricle size. Our second claim is that a scale-normalized margin readout will converge to edits whose deformation pattern matches what classical morphometry and microstructure imaging independently find in the same infants. Those edits should shrink the thalamus and lentiform nucleus, change white-matter contrast in the subplate-transition zones, and alter posterior cortical microstructure.
The closest existing work is ICAM-Reg. It translates dHCP neonatal scans along a learned latent age attribute and reads off attribution maps from decoder differences. It found punctate white-matter lesions without lesion labels. It trains its regressor with a smooth L1 loss and a GAN discriminator, never examines whether its search or discriminator saturates, and produces a map rather than a calibrated distance. We differ in three ways. Our edits live in an explicit, interpretable space of deformation plus low-frequency contrast change, rather than a decoder that can hallucinate. We treat the objective's saturation as the variable under study. And we validate maps against non-learned biomarkers in the same subjects.
Research Program
The maturity model is a 3D convolutional network trained on dHCP T2-weighted and T1-weighted scans spanning 24–45 weeks PMA, including preterm infants scanned soon after birth. We give it an ordinal head: for each week threshold $k$, a logit $s_k(x)$ scores "PMA exceeds $k$ weeks." The same backbone can also carry a 22-way softmax head over weekly bins and a scalar regression head. All three are trained on identical data, so only the readout differs.
For a preterm infant's term-equivalent scan $x$, we search over edits of the form $x' = (x \circ \phi) + b$. Here $\phi$ is a smooth diffeomorphic deformation and $b$ is a spatially smooth intensity field. We include $b$ because white-matter maturation shows up as a change in T2w/T1w contrast, not only as a change in shape, as Nazeri et al. showed. The maturational distance toward target age $a^*$, which is 40 weeks by default, is
$$D(x) = \min_{\phi,\, b}\; \|\log J_\phi\|_1 + \lambda \|b\|_2 \quad \text{subject to the model reading } x' \text{ as} \ge a^*,$$
where $J_\phi$ is the voxelwise Jacobian determinant of the deformation, which measures local volume change, and $\lambda$ trades shape against contrast. We solve this with a penalty method. The objective that drives the model's reading is the variable we manipulate.
The cross-entropy version maximizes the softmax probability of the target bin. The margin version, adapting the Carlini–Wagner clipped margin and the scale normalization of DLR from AutoAttack to ordinal thresholds, is
$$\mathcal{L}_{\text{margin}}(x') = \frac{\sum_k \max\big(-y_k\, s_k(x'),\, -\kappa\big)}{\max_k s_k(x) - \min_k s_k(x)}.$$
In this expression, $y_k = +1$ for thresholds below $a^*$ and $-1$ otherwise. The constant $\kappa$ stops the loss from rewarding over-shooting. The denominator, computed once on the unedited scan and held fixed, removes dependence on the model's arbitrary logit scale. That last point matters for comparing distances across models and across scanners, a known weak spot in neonatal imaging: the Vogt-Bailey cortical homogeneity findings did not fully replicate between TEBC and dHCP, per the feature-similarity gradients study. Losses are computed per subject and summed, never averaged over a batch, because Redgrave et al. showed that batch-mean scaling alone can drive gradients to underflow and fake robustness.
The adversarial-robustness papers supply the method; none has touched regression, ordinal outputs, or 3D medical volumes. The neonatal papers supply the validation targets, which are biology measured without any neural network. Boardman et al.'s deformation-based morphometry found thalamic and lentiform contraction at term-equivalent age, worse below 28 weeks gestation (t = 7.13), with no ADC difference. That pattern indicates disrupted growth rather than tissue destruction, and it is exactly the shape change a deformation edit should reproduce. Nazeri et al. found elevated T2w/T1w ratios in preterm white matter, concentrated in the intermediate and superficial patterns that track subplate dissolution, and the contrast field $b$ should reproduce those. Dimitrova et al. found posterior cortical diffusivity and neurite-density deviations at term-equivalent age, and it also supplies the constraining negative result: those deviations did not predict Bayley-III scores at 18 months. We do not assume our distance will overturn that negative result. We test it.
Baselines
The naive reference point is the brain-age gap from the scalar regression head trained with mean squared error. We take the predicted PMA minus the true PMA at scan and apply the standard linear correction for regression toward the mean, fitted on term infants. This baseline has no search and no saturation problem in its training loss, so it isolates what the counterfactual machinery adds. Beating it means two things. A higher within-group Spearman correlation with the deep grey matter Jacobian contraction among infants born before 28 weeks, where point predictions compress most. And a higher correlation with gestational age at birth among infants scanned at term-equivalent age.
The strongest existing individual-mapping method is ICAM-Reg, retrained on the same dHCP split with its published architecture and given the same hyperparameter-tuning budget. We read off its attribute-translation difference maps toward 40 weeks. The file contains no reported dHCP birth-age MAE for it. The only figure on record is an MAE of 2.20 ± 1.86 years for brain age on UK Biobank in the ICAM-Reg paper, which is not comparable to a neonatal task. The adult analogue for the ranking-loss alternative is TSAN's MAE of 2.428 years on OASIS/ADNI/PAC-2019 in the cascade-network paper. No reported result was found on Lacuna for neonatal PMA prediction or individual delay mapping on dHCP, so we will establish these numbers ourselves. Beating ICAM-Reg means a higher voxelwise spatial correlation between our edit maps and the independent biomarker maps (DBM Jacobians, T2w/T1w deviation, Vogt-Bailey index), averaged per subject and tested by paired permutation.
The non-learned reference is Gaussian process normative modelling in the style of Dimitrova et al., fitted on dHCP term infants: per-region Z-scores for diffusion metrics and thickness, plus DBM Jacobians. These are the scores our method must at least match in predicting 18-month outcomes.
The key ablations hold everything fixed except one ingredient at a time: the backbone, edit space, regularizer, optimizer, step budget, and random restarts. We compare cross-entropy against margin on the softmax/ordinal heads. We compare margin with and without the scale normalization. We compare per-subject against batch-mean loss. And we compare 20 against 300 optimization steps. That last ablation replicates the finding of the transferable targeted attacks paper that extra iterations help only when the loss does not saturate.
Why not just…?
Why not just use a regression head with squared error and bias correction? This is the most serious threat. A squared-error gradient is proportional to the residual, so it does not saturate in the softmax sense, and a counterfactual search targeting "predicted PMA = 40" through a regression head may well work. We include exactly that search as an arm of the ablation. If it matches the margin search on biomarker agreement and distance calibration, the honest conclusion is narrower. It would be "never use cross-entropy or binned-softmax readouts for counterfactual delay estimation," which still matters because binned soft-label heads are common in brain-age work. Two regimes are where we expect regression to fall short, and we will check both. One is extreme immaturity, where predictions compress toward the training mean. The other is cross-scanner comparison, where the raw regression output has no built-in scale normalization.
Why not just run more optimization steps with the standard loss? Evidence from natural images says this does not help once the loss saturates: in the transferable targeted attacks paper, cross-entropy searches plateaued while logit-loss searches kept improving and reached 99.4% targeted transfer. Our step-count ablation tests whether the same holds for 3D neonatal volumes. If cross-entropy distances keep growing with steps and converge to the margin distances, the saturation hypothesis is wrong for this setting and we will report that.
Why not just use deformation-based morphometry or normative Z-scores and skip the network? These methods already find the group effects, and they are our validation targets precisely because they do not depend on a learned model. Their limits are specific. DBM measures volume change against a single term template and says nothing about contrast-based white-matter immaturity. GPR Z-scores are univariate per metric. Neither yields a single individual distance that combines shape and contrast. If our distance adds nothing over a simple sum of a subject's absolute Z-scores in predicting gestational age at birth or 18-month Bayley-III scores, the network is not earning its place. That comparison is part of the experiments below.
Why not just use a linear latent counterfactual model? The latent 3D counterfactual model avoids gradient search with a closed-form linear model in VQ-VAE latent space. Its authors note the linearity limitation, and it has been tested only on adolescents and older adults. We include it as a secondary comparison on the longitudinal check below. We expect it to miss region-specific timing, because white-matter maturation proceeds in a deep-to-superficial wave in which different regions change at different times.
Proposed Experiments and Decisive Checks
The primary data are dHCP structural and NODDI diffusion scans, with TEBC as an external cohort. Training and validation splits are made by subject.
The first decisive check tests whether the search optimizer, rather than the brain, is determining the measured delay. For every preterm infant at term-equivalent age, we compute distances under each objective and record the fraction of searches that converge within budget. We also measure how the distance changes between 20 and 300 steps. The saturation hypothesis predicts that cross-entropy searches fail or stall more often and give smaller distances, disproportionately in infants born before 28 weeks.
The second check tests agreement with biology. Per-subject edit Jacobians are correlated with that subject's DBM Jacobians in the thalamus and lentiform nucleus. The contrast field $b$ is correlated with T2w/T1w deviation in the intermediate and superficial white-matter patterns and with the Vogt-Bailey index in posterior cortex. The margin objective should win on these correlations, and the gain should be largest in the most premature infants.
The third check is ground truth from real development, run where dHCP contains preterm infants scanned twice, once early and once at term-equivalent age. We edit the early scan toward the PMA of the later scan and correlate the edit's Jacobian with the actual early-to-late registration Jacobian. This is the only check in which the correct answer is observed rather than inferred, so it is the result we would weight most.
The fourth check is transfer. We apply the dHCP-trained model to TEBC without retraining. We compare the correlation between distance and gestational age at birth with and without scale normalization. If normalization helps across scanners, the shift-and-scale invariance argument from adversarial evaluation carries over.
The fifth check is functional, and it is deliberately pre-registered as a test that may come out null. Using dHCP 18-month Bayley-III scores, we ask whether maturational distance explains outcome variance beyond gestational age and the Gaussian process Z-score summary. A null result would be consistent with Dimitrova et al. and would not count against the measurement claims. A positive result would be the first individual learned delay measure to break that dissociation.
References
- ICAM-Reg: Interpretable Classification and Regression With Feature Attribution for Mapping Neurological Phenotypes in Individual Scans
- Neurodevelopmental Patterns of Early Postnatal White Matter Maturation Represent Distinct Underlying Microstructure and Histology
- Towards Evaluating the Robustness of Neural Networks
- Reliable evaluation of adversarial robustness with an ensemble of diverse parameter-free attacks
- Feature similarity gradients detect alterations in the neonatal cortex associated with preterm birth
- Generating Adversarial Samples in Mini-Batches May Be Detrimental To Adversarial Robustness
- Abnormal deep grey matter development following preterm birth detected using deformation-based morphometry
- Preterm birth alters the development of cortical microstructure and morphology at term-equivalent age
- Brain Age Estimation From MRI Using Cascade Networks With Ranking Loss
- On Success and Simplicity: A Second Look at Transferable Targeted Attacks
- Latent 3D Brain MRI Counterfactual