Brain age models predict a person's age from an MRI scan. The gap between predicted and actual age is then read as a biomarker: a positive gap suggests the brain looks older than it should. These models inherit a basic confusion from their training data. A single scan mixes two things. One is baseline anatomy, meaning head size, cortical shape, and ventricle configuration, which differ across people and across sex and ancestry groups for reasons unrelated to aging. The other is atrophy, the tissue loss that accumulates over a lifetime. A network trained on mostly White European cohorts learns that certain baseline shapes co-occur with certain ages. When it sees a person from an underrepresented group whose baseline anatomy differs, it reads that difference as extra or missing years. We propose to build a brain age model that cannot make this mistake by construction. The model explains each scan as a subject-specific baseline anatomy deformed by a physically modeled atrophy process, and it lets age be read only from the atrophy.
The mechanism combines two tools. Deep unfolding turns an iterative optimization algorithm into a fixed-depth trainable network. We use it to recover the atrophy field from the scan. Recovering that field requires solving a tissue-deformation equation many times, once per network stage per training example, and this is where computational cost becomes the binding constraint. MRI tissue is a high-contrast mixture of gray matter, white matter, and cerebrospinal fluid, which is exactly the setting where fine-grid solvers are expensive and multiscale basis methods pay off. We make the approach affordable by predicting multiscale basis functions with a neural operator, following Locally Subspace-Informed Neural Operators for Efficient Multiscale PDE Solving. The claim to test is that a physically constrained split between baseline and atrophy removes most of the race and sex disparity in brain age error. It should do so without using demographic labels during training and without giving up clinical sensitivity.
Core Claim and What Motivates It
The strongest evidence for the problem comes from Analysing race and sex bias in brain age prediction. That study trained a 3D ResNet-34 on Cam-CAN and IXI (IXI contains 451 White, 14 Black, and 50 Asian subjects) and tested it on balanced UK Biobank subgroups of 126 subjects each. It found significant error differences between Black and White subjects (p=0.022), Black and Asian subjects (p=0.013), and male and female subjects (p=0.008). Its penultimate-layer analysis shows where the bias lives. The first three principal components tracked age, the fourth separated racial groups, and 7 of 12 pairwise feature-distribution comparisons differed. So demographic structure sits inside the representation that feeds the age readout.
Our hypothesis is that much of this structure is baseline anatomy that the network has no reason to discard. If that is right, the fix is architectural. The model should give age only a channel that baseline anatomy cannot enter. The alternative fix, scrubbing demographics out of the features, removes real biology along with the artifact.
Two further findings constrain the design. Benchmarking the generalizability of brain age models shows that post-hoc linear corrections can improve accuracy numbers while distorting which regions drive predictions. The same paper shows that the most accurate model (gradient tree boosting on extracted features) had less cognitive utility than a less accurate CNN. We therefore treat clinical sensitivity as a co-equal outcome, not an afterthought. Neurofind, a normative outlier model built on healthy controls, cautions that normative models fail when the reference data lacks diversity. Our design aims to reduce how much diversity the reference data needs.
Research Program
We model a T1-weighted scan $I$ as a baseline anatomy $B$ warped by a displacement field $u$. The displacement solves a heterogeneous elasticity-type equation driven by a scalar atrophy field $a(x)$, which measures local tissue loss:
$$-\nabla\cdot\big(\kappa(x)\,\nabla u\big) = \nabla a, \qquad I \approx B\circ(\mathrm{id}+u).$$
Here $\kappa(x)$ is a high-contrast tissue stiffness map taken from the segmentation. This equation is the standard way biophysical atrophy simulators turn tissue loss into shape change. Its key property for us is that a smooth global rescaling or reshaping of the head cannot be produced by any plausible $a$. Such variation is forced into $B$.
Estimating $(B, a)$ from $I$ is an inverse problem, and we unroll an alternating proximal-gradient solver for it over $K$ stages. The unrolling follows the structure introduced by ISTA-Net, which turns the Iterative Shrinkage-Thresholding Algorithm into a network with learned step sizes and learned convolutional proximal modules. It reported 32.57 dB on Set11 at a 25% sampling ratio, with GPU reconstruction about 1,000 times faster than the iterative solver it replaces. We borrow that structure but change the forward operator. In our network the data-consistency step differentiates through the atrophy equation above. A learned prior step updates $B$, and a separate learned prior step updates $a$.
Two training signals shape the channels. First, a small regressor maps the recovered atrophy field to age on healthy training subjects only, which is the usual normative setup. Second, the baseline channel is penalized adversarially for predicting age. This asks the network to push age information out of $B$ using only age labels. Race and sex labels are never used in training; they are reserved for evaluation. This matters because the groups most affected are the ones with the fewest labeled examples. The age gap is the atrophy-based prediction minus chronological age. Because $a(x)$ is a spatial map, regional gap maps come for free, which addresses the loss of regional specificity documented by Local Brain-Age: A U-Net Model.
The cost problem is concrete. At 1 mm resolution each forward and adjoint solve involves on the order of millions of unknowns. With $K$ unrolled stages and backpropagation, every training example needs roughly $2K$ solves, and unrolled networks already store activations at each stage, which limits depth. We replace the fine-grid solve with a coarse multiscale solve. The locally adapted basis functions come from a factorized Fourier neural operator fed with the subject's $\kappa$, trained with the subspace-alignment loss of the GMsFEM-NO paper cited above. That paper reports 60 to 100 times faster basis generation with 1–3% solution error on stochastic high-contrast permeability benchmarks. It also reports that standalone neural operators blew past 100% error under out-of-distribution forcing, while the subspace-informed version stayed stable.
MRI voxel grids are Cartesian, so the structured-grid restriction that paper lists as a limitation does not bite here. The open question is whether basis accuracy holds on real tissue maps rather than synthetic permeability fields. The classical construction, GMsFEM with the cheaper randomized snapshots of Randomized Oversampling for GMsFEM, gives us a reference to check against.
This proposal differs from prior brain age architectures in what it leaves out of the age pathway. The graph networks of Geometric deep learning on brain shape predict age from cortical meshes (MAE 4.58 years) and sex (88% accuracy) from the same representation. The Global-Local Transformer reaches 2.70 years MAE. The covariance networks of Transferable coVariance Neural Networks transfer across atlas resolutions. All of them let baseline shape and aging share one pathway into the age readout. None uses a physical model, and none evaluates racial fairness except the ResNet study, which diagnoses the problem without proposing a fix.
Baselines
The naive reference point is a model that predicts each subject's age as the training-set mean, alongside chronological-age-only normative residuals. These anchor what any MAE means and give a floor for clinical effect sizes.
The primary competitor is a direct reproduction of the 3D ResNet-34 from Analysing race and sex bias in brain age prediction. It is trained on the same Cam-CAN and IXI data and tested on the same six balanced UK Biobank subgroups of 126 each. It is the strongest competitor on this specific question because it defines the documented disparity. The file reports significance levels (p=0.022 Black vs White, p=0.013 Black vs Asian, p=0.008 male vs female) but not per-subgroup MAE values. No reported result was found on Lacuna for subgroup-level MAE on UK Biobank, so we regenerate those numbers ourselves.
As a stronger accuracy baseline we include the Global-Local Transformer. On its own 8,379-subject multi-site cohort it reported 2.70 years MAE, against 3.04 for the simple fully convolutional network (SFCN) and 3.15 for ResNet101 Global-Local Transformer. We retrain it on our training data rather than quote that number, since the cohorts differ.
We also include three alternative remedies. The first is gradient tree boosting on FreeSurfer features, the accurate model in Benchmarking the generalizability of brain age models. The second is ResNet-34 plus post-hoc linear correction for age, sex, and ethnicity. The third is ResNet-34 with adversarial removal of race and sex from its features plus group-reweighted loss; unlike our method, this baseline receives demographic labels.
Fairness of comparison is enforced in four ways. All models see identical preprocessing and identical training subjects. All get the same hyperparameter search budget, measured in GPU-hours and number of trials, tuned on a held-out Cam-CAN/IXI validation split, never on UK Biobank. Model selection uses overall validation MAE only. For compute, we report our wall-clock training cost against the ResNet's.
Three ablations isolate the key ingredient. The first keeps the unrolled baseline-plus-deformation network but replaces the atrophy equation with a free, learned smooth deformation. This tests whether the physics or just the two-channel split does the work. The second runs the same inverse problem with a classical non-unrolled iterative solver on the fine grid, measuring the accuracy and cost that unrolling buys. The third swaps neural-operator bases for fine-grid solves and for classical randomized GMsFEM bases, measuring what multiscale acceleration costs in gap accuracy.
Beating the baselines means three things. Subgroup MAE disparities must shrink, with bootstrapped confidence intervals on the Black-versus-White and male-versus-female MAE differences excluding the ResNet-34 disparity. Overall MAE must stay within 0.5 years of the ResNet-34. Clinical effect sizes must be no worse.
Why not just…?
Why not just correct the output with a post-hoc linear adjustment for demographics? This shifts the scalar prediction but leaves the leaked structure in the features. The fourth principal component still encodes race Analysing race and sex bias in brain age prediction. Such corrections can also produce artificial accuracy gains that distort regional interpretation Benchmarking the generalizability of brain age models. The correction also needs reliable per-group coefficients, which cannot be estimated from 14 Black subjects. It remains a baseline. If it closes the gap as well as our model and matches it on clinical effect sizes, the physics buys nothing for fairness, and we would say so.
Why not just train a CNN with adversarial debiasing or group-balanced reweighting? This is the real threat. Our experiments settle it directly by giving that baseline demographic labels our model never sees. Two considerations favor our approach, and both are testable. First, forcing features to be race-invariant also erases any genuine group differences in aging rate, such as vascular contributions. Our model instead allows demographic differences in atrophy and removes only baseline shape. We check this by testing whether our gaps preserve known associations with clinical status while the adversarial model flattens them. Second, adversarial training depends on labels that are scarce in exactly the affected groups. If the adversarial model matches ours on disparity, accuracy, and clinical effect size, the contribution narrows to label-free debiasing and interpretable regional maps.
Why not just use tabular morphometry with gradient tree boosting, or registration-derived deformation features, and skip the unrolled PDE? Tabular models were more robust to scanner changes but missed cognitive differences that a CNN captured Benchmarking the generalizability of brain age models. Global morphometry also washes out regional signal that voxel-level models detect Local Brain-Age. Deformation features from registration to a single template inherit that template's shape, which is the source of the bias we target. The free-deformation ablation tests whether learning the split end-to-end is enough without physics. The classical-solver ablation tests whether unrolling is needed at all, or only makes things cheaper.
Proposed Experiments and Decisive Checks
The central experiment trains on Cam-CAN (n=652) and IXI (n=563) and tests on the six balanced UK Biobank subgroups, mirroring the published protocol.
We report per-group MAE and the rank-based group comparisons used in the original study (Kruskal-Wallis with Conover-Iman follow-up tests). A non-significant result at n=126 is weak evidence, so we add bootstrapped confidence intervals and equivalence tests on the MAE differences. We also report results on the full 42,786-subject UK Biobank test set for power.
Representation leakage is measured by repeating the principal-component and Kolmogorov-Smirnov analysis on the atrophy channel and the baseline channel separately. The prediction is that race separability concentrates in $B$ and falls in $a$. We quantify this with linear-probe accuracy for race and sex on each channel.
Clinical utility is tested on OASIS3, comparing progressive and stable mild cognitive impairment. We compare against the subcortical effect sizes above 1.5 reported by Local Brain-Age, and we test within-subject repeat scans to check that the atrophy channel is stable between sessions and grows between years. Scanner robustness is checked on the PDS and HCP-Development cohorts used in Benchmarking the generalizability of brain age models, which were acquired on Siemens Trio and Prisma scanners. A physically constrained atrophy channel should absorb less scanner offset than a black-box model, since contrast changes are not expressible as tissue loss.
The cost study reports solves per training step, memory, and wall-clock time for fine-grid, classical multiscale, and neural-operator multiscale variants. The target is that the neural-operator version makes $K\ge 8$ stages trainable on a single GPU node without the gap estimates drifting by more than a small fraction of a year relative to fine-grid solves.
The decisive failure mode is identifiability. The baseline prior might absorb large ventricles and widened sulci as normal anatomy, emptying the atrophy channel of age information. We detect this early on synthetic scans. These are built by applying known atrophy fields to young Cam-CAN brains of varied baseline shape, and the test is whether recovered $a$ matches ground truth while $B$ stays age-free. If recovery fails there, the physical split is not doing its job, and we would report that before any fairness claim.
References
- Locally Subspace-Informed Neural Operators for Efficient Multiscale PDE Solving
- Analysing Race and Sex Bias in Brain Age Prediction
- Benchmarking the generalizability of brain age models: Challenges posed by scanner variance and prediction bias
- Neurofind: using deep learning to make individualised inferences in brain-based disorders
- ISTA-Net: Interpretable Optimization-Inspired Deep Network for Image Compressive Sensing
- Local Brain-Age: A U-Net Model
- Generalized multiscale finite element methods (GMsFEM)
- Randomized Oversampling for Generalized Multiscale Finite Element Methods
- Geometric deep learning on brain shape predicts sex and age
- Global-Local Transformer for Brain Age Estimation
- Predicting Brain Age Using Transferable Covariance Neural Networks