When a teenager's cortex thins, or an older adult's ventricles enlarge, the scan alone cannot say whether this is disease or ordinary growth and aging. Current "normative" models compare each person with a healthy reference population, but they compare states: the brain's intensity or regional volume at one moment, after warping it onto a single fixed template brain. This proposal compares motion instead. For each person we estimate the smooth, fold-free deformation that carries their earlier scan onto their later one, a "velocity field" saying how each bit of tissue moved per year. We then ask whether that motion is unusual relative to healthy people at the same stage of life. That comparison is only valid if every person's motion is expressed in a common coordinate system that itself changes with age, rather than one static template.
The opportunity is that each ingredient exists separately but has never been combined. Longitudinal diffeomorphic registration (alignment that is smooth and invertible, so tissue never tears or folds) can now be learned efficiently. Normative deviation scoring is mature. Population atlases built from deformations are well understood. What is missing is a normative distribution over transported longitudinal velocity fields, conditioned on age. The point is to subtract the expected maturational or aging motion and score the residual motion as candidate pathology.
Core Claim
The claim has two parts. First, a person's longitudinal deformation, once moved into an age-matched template and compared against the healthy deformation expected at that age, separates pathology from maturation better than comparing single scans against a normative model. Second, the age-matched coordinate system is the essential ingredient. Doing the same comparison in one static template should leave a measurable age-dependent bias in healthy people, because the template warp distorts tissue more at the age extremes.
Concretely, let $v_i$ be subject $i$'s per-year velocity field after transport into the template $T(a_i)$ at their mean age $a_i$. Let $\mu(a)$ and $\Sigma(a)$ be the healthy mean field and a low-rank-plus-diagonal covariance at age $a$. The voxelwise residual and its deviation score are
$$r_i(x) = v_i(x) - \mu(a_i)(x), \qquad z_i(x) = \sqrt{r_i(x)^\top \Sigma(a_i)(x)^{-1} r_i(x)},$$
where $x$ indexes voxels. This writes the gap that the research file attributes to TimeFlow as an explicit decomposition: observed motion equals expected normative motion plus a residual. The residual is what we score.
Relation to Prior Work
The motivating evidence comes from the developmental and aging literature. Longitudinal Correlation Analysis shows that within-subject differences cancel static individual traits such as head size and genetics. Its learned representations track age at $r \geq 0.4$, against $r \leq 0.15$ for cross-sectional methods, and they recover the post-20 developmental plateau. However, it subtracts intensities after static normalization, so a boundary that moves between visits appears as an intensity change. We keep its within-subject logic but replace subtraction with an estimated deformation. Ehsan Adeli et al. supply the same motivation about cross-sectional noise. Kerstin Ritter et al. motivate the aging side: normal atrophy and Alzheimer's atrophy are hard to separate.
The normative-modeling methods we compete with are mmNormVAE and mmSIVAE. mmNormVAE scores voxelwise reconstruction error of single scans, using a model trained on 9,633 UK Biobank controls. mmSIVAE scores latent Mahalanobis distance over 90 FreeSurfer regions. Both assume a static, already-normalized brain and evaluate one timepoint at a time. Our proposal scores change, voxelwise, in age-matched coordinates. DIVE shows that sub-regional resolution reveals patterns that regional averages hide, such as typical Alzheimer's versus posterior cortical atrophy. But it models scalar thickness on fixed vertices with sigmoid trajectories, which cannot represent developmental plateaus. We inherit its argument for spatial resolution but not its parametric curve.
The deformation machinery is reused rather than invented. TimeFlow gives time-conditioned longitudinal registration with a diffeomorphic variant built on stationary velocity fields. We use it for within-subject velocities, and in a second instance conditioned on age for the template trajectory. TimeFlow itself builds no population norm and no cross-subject coordinate system, and that is exactly what we add. Bayesian Mixed Effect Atlas Estimation supplies the idea of treating the template and subject deformations as jointly estimated, topology-preserving random variables. It is static and was tested only on small shapes, so we extend it along age.
The constraining evidence comes from three sources. The newborn connectome study found static atlases distort rapidly growing brains badly enough that its authors abandoned registration altogether. We take this as the failure mode our age-matched template must fix, without giving up cross-subject correspondence. CRONOS shows that in slowly changing anatomy, carrying the last scan forward is a strong baseline for image-fidelity metrics. On Lumiere glioma MRI it scores an NRMSE (normalized root-mean-square error) of 8.38, against CRONOS's 7.55. We therefore do not judge success by reconstruction metrics at all; we judge it on clinical association and confound removal. CRONOS also does not enforce topology, which is one reason we use diffeomorphic fields. The surgical-video deformation methods Forplane and Endo-4DGS are fast, but they solve a rendering-speed problem over seconds. Registration of a handful of MRI volumes is not bottlenecked there, so we deliberately do not adopt them.
Research Program
The pipeline has four stages.
Stage 1: within-subject velocities. For each subject with two or more scans, we estimate a diffeomorphic velocity field in their own native space using TimeFlow's diffeomorphic variant. We normalize it by the inter-scan interval to get motion per year. Irregular gaps are handled through TimeFlow's continuous time conditioning.
Stage 2: age-conditioned template. We fit a template trajectory $T(a)$ from healthy controls. A cross-sectional age-conditioned atlas network produces a template at any age, linked across ages by a diffeomorphic flow. The mixed-effect formulation is used so that the template is not biased toward any one age bin.
Stage 3: transport. Each subject's velocity field is mapped into $T(a_i)$ by pushing it forward through the subject-to-template diffeomorphism. Pushing forward rotates and scales the vectors by the local Jacobian; it does not merely resample them. For pooling across ages, fields are then carried along the template trajectory to a reference age.
Stage 4: normative model and calibration. We fit $\mu(a)$ and $\Sigma(a)$ on healthy fields using smooth age dependence. Because the file flags Gaussian deviation scores as unreliable for these distributions, we convert $z_i(x)$ to calibrated tail probabilities using held-out healthy subjects. We do not trust the Gaussian form.
Two ablations isolate the claim. The first keeps everything but uses one static template, which tests whether age-matched coordinates matter. The second keeps everything but replaces the full vector field with its scalar log-Jacobian, the local volume change. This tests whether motion direction carries information beyond volume change.
Baselines
Naive reference. The do-nothing comparator is classic longitudinal tensor-based morphometry. We register all scans to one static template and compute each subject's per-year log-Jacobian of within-subject change. We then regress it on a polynomial in age in healthy controls and score patients by voxelwise residual Z-scores. It shares our registrations, splits and controls, so any gain from our method must come from transport, age-matched coordinates or direction information. It is also the strongest simple alternative, discussed below.
Strongest existing method. This is mmNormVAE, re-trained under its published recipe on our healthy training set and applied to the patients' scans. We report its score both at the baseline visit and as a change between visits, so it is not handicapped by being cross-sectional. On 862 ADNI patients it reports a Pearson correlation between mean deviation and ADAS13 cognitive score of $r = 0.542$ (multimodal), $r = -0.494$ for RAVLT, and a deviation slope across disease stages of 0.2. Those figures come from its own cohort and will not be directly comparable to ours. Beating it means a higher ADAS13 association on the identical ADNI subjects and splits. It also means a significantly better area under the curve for MCI-to-Alzheimer's conversion, with paired bootstrap confidence intervals.
Within-subject intensity baseline. We include an LCA-style model that applies normative scoring to voxel differences in static template space. This isolates whether a deformation is better than subtraction. Its NCANDA age correlation of $r \geq 0.4$ is the reference for the developmental arm.
Regional baseline. We fit an mmSIVAE-style normative model on FreeSurfer regional volume change rates. No numeric result for mmSIVAE is reported on Lacuna in the file.
Tuning budget. Every model receives the same hyperparameter search budget on a validation split of healthy subjects only. Patients never touch tuning.
Why not just…?
Why not just use log-Jacobian tensor-based morphometry with an age covariate? This is the real threat. It captures most atrophy signal, and if our method only matches it, the extra machinery is unjustified. Two experiments settle it. The direction ablation tests whether full vectors add information on the same subjects. The second test is the healthy-control age-bias check described below. A static template should produce higher false-positive rates at the age extremes than the age-matched one. If both tests show no difference, we will report that the simple method suffices and restrict our claim accordingly.
Why not just run mmNormVAE on each visit and difference the scores? Differencing two state deviations still compares each scan against a static template. It also inherits the static inter-individual noise that LCA shows overwhelms developmental signal, where cross-sectional age correlation was $\leq 0.15$. We run exactly this variant as part of the strongest baseline, so the comparison is direct.
Why not just model regional volume change rates? DIVE found that regional averaging hides spatially distinct atrophy patterns. The regional baseline quantifies how much is lost on our cohorts. If it matches us on clinical association, the voxelwise result should still matter for spatial specificity. We test that on DRC-style phenotype separation where such data are available, and otherwise report the regional baseline as sufficient for staging.
Proposed Experiments
Aging arm (ADNI and OASIS). We use ADNI cognitively normal subjects with at least two scans to fit the norm, holding out a fraction of them. Patients across subjective memory complaints, early MCI, late MCI and Alzheimer's disease are scored. Outcomes are correlation with ADAS13, RAVLT and MMSE, separation between disease stages, and conversion prediction. OASIS serves as a zero-shot test across scanners, following TimeFlow's use of it, with no re-fitting of the norm.
Developmental arm (NCANDA). The norm is fit on healthy youths aged 12–25. Here success means three things. Calibrated deviation scores of held-out healthy youths must be flat across age. The fitted normative speed $\|\mu(a)\|$ must reproduce the plateau after age 20 that LCA reports. And the residual must be uncorrelated with age while still correlating with available behavioral measures.
Decisive check. The single most decisive test is the healthy-control age-bias test, run in both arms. We plot the false-positive rate at a fixed calibrated threshold as a function of age, under the static-template ablation and under the full method. The proposal's central claim predicts a U-shaped or monotone age dependence for the static template that largely disappears with age-matched coordinates. Failing this check would falsify the spatial-normalization half of the claim, regardless of the clinical correlations.
Sanity checks. Two further checks guard against artifacts. Using fewer than about one percent of voxels with a non-positive Jacobian guards against folding introduced in transport. Scan-rescan pairs with near-zero interval should yield near-zero velocity and deviation, which guards against the registration noise floor being mistaken for signal.
References
- Longitudinal Correlation Analysis for Decoding Multi-modal Brain Development
- Pioneering new paths: the role of generative modelling in neurological disease research
- mmNormVAE: Normative Modeling on Multimodal Neuroimaging Data using Variational Autoencoders
- Multimodal normative modeling in Alzheimers Disease with introspective variational autoencoders
- DIVE: A spatiotemporal progression model of brain pathology in neurodegenerative disorders
- Temporal Conditioning for Longitudinal Brain MRI Registration and Aging Analysis
- Bayesian Mixed Effect Atlas Estimation with a Diffeomorphic Deformation Model
- Atlas-independent brain connectome analysis at voxel-level granularity: graph convolutional networks for etiology classification in newborns
- CRONOS: Continuous time reconstruction for 4D medical longitudinal series
- Efficient Deformable Tissue Reconstruction via Orthogonal Neural Plane
- Endo-4DGS: Endoscopic Monocular Scene Reconstruction with 4D Gaussian Splatting