Surface tensor-based morphometry describes how each small piece of a brain structure's surface is stretched, sheared, or compressed relative to a template. On the hippocampus it is among the most sensitive structural markers of Alzheimer's disease and amyloid burden. The measurements live on roughly 30,000 surface vertices, which is far more than the number of subjects, so every published pipeline first groups vertices into patches and works with patch summaries. The patch layout is chosen by hand. Geographers have a name for what this can do to results: the modifiable areal unit problem. When a continuous field is summed into containers, changing the container size (the scale effect) or shifting the boundaries (the zoning effect) can change correlations, the features a model selects, and even the sign of an effect. Nobody has checked whether the celebrated hippocampal findings, such as the early CA1 and subiculum involvement, the effect-size cutoffs, and the cross-cohort accuracies, survive a change of patch grid.
This note proposes to test that, and to argue that the question becomes urgent once data from several scanners and cohorts are pooled. Scanner effects are not uniform over the surface. They are spatially structured distortions of the tensor field, so the size of a "site effect" estimated on patches depends on how the patches are drawn. Our core claim is twofold. First, aggregating patches before removing site effects makes the harmonization depend on the chosen partition, and this dependence inflates cross-cohort instability. Second, removing site effects at the vertex level, in a geometry that respects the tensors, and then averaging conclusions over a distribution of random patch partitions yields biomarker maps and classifiers that are both more stable across partitions and transfer better across cohorts.
Core Claim and Mechanism
Each vertex $v$ carries a deformation tensor $T_v$, a symmetric positive-definite (SPD) matrix. SPD matrices form a curved space, so ordinary averaging distorts them. The log-Euclidean approach, used in Multivariate Statistics of Tensor-Based Cortical Surface Morphometry and as the tractability device in Riemannian Nonlinear Mixed Effects Models, takes the matrix logarithm $L_v=\log T_v$ and works linearly on $L_v$. A patch $P$ is then summarized by $\bar L_P = |P|^{-1}\sum_{v\in P} L_v$.
Suppose site $s$ acts on subject $i$ at each vertex through a shift and a scale, in the style of ComBat:
$$L_{v,i} = \mu_v + \beta_v x_i + \gamma_{v,s} + \delta_{v,s}\,\epsilon_{v,i}.$$
Here $\mu_v$ is the template mean, $x_i$ the subject's covariates with effects $\beta_v$, $\gamma_{v,s}$ the additive site shift, $\delta_{v,s}$ the multiplicative site scale, and $\epsilon_{v,i}$ the residual. The shift $\gamma$ commutes with patch averaging: the patch shift is just the mean of the vertex shifts. The scale does not commute. The patch residual variance at site $s$ is
$$\mathrm{Var}(\bar L_P \mid s) = |P|^{-2}\sum_{u,v\in P}\delta_{u,s}\delta_{v,s}\,\mathrm{Cov}(\epsilon_u,\epsilon_v).$$
This quantity depends on which vertices share a patch and on their residual correlation. A scale factor fitted to patch summaries, as CB-Predict does on fixed regions in ComBat-Predict enhances generalizability of neuroimaging models to new sites, therefore changes when the partition changes. The size of that change is predictable: it grows with the spread of $\delta_{v,s}$ inside a patch and with the residual correlation. That gives a falsifiable mechanism. Partition sensitivity of harmonized results should be largest in regions where site scale varies sharply across the surface, plausibly near segmentation boundaries where registration errors concentrate. A vertex-level correction, by contrast, defines a single harmonized field that any partition can then summarize consistently.
Research Program
The program has three components.
The first component is a partition-sensitivity audit. We generate random partitions of the hippocampal mesh by geodesic Voronoi tessellation. Scale is controlled by the number of seeds, and zoning is varied by redrawing seeds at a fixed count. This brings the "space as a distribution of partitions" idea of SPAD, from Fairness by "Where", onto a curved surface; SPAD itself names non-uniform geometries as unsolved. SPAD uses partitions to equalize regional performance. We use them for a different job: measuring how much scientific conclusions move.
We rerun three published pipelines under hundreds of partitions:
- Patch selection by Hotelling's $T^2$ with the 0.8 Mahalanobis effect-size cutoff, from Classification of Alzheimer's disease based on hippocampal multivariate morphometry statistics.
- PASCS-MP, patch sparse coding with a correntropy loss followed by max-pooling, from Predicting Brain Amyloid Using Multivariate Morphometry Statistics, Sparse Coding, and Correntropy.
- Patch group-lasso selection, from Federated Morphometry Feature Selection.
For each pipeline we record three outcomes: the vertex-wise selection frequency, the variance of the effect size, and the spread of accuracy across partitions. We do this before and after harmonization. All three papers fix one patch layout, which is exactly the gap.
The second component is the harmonize-before-aggregate estimator. We fit the model above in log-Euclidean coordinates; for 2×2 surface tensors these are the three free entries of the symmetric matrix $L_v$. Two choices keep the estimate stable with few subjects per vertex. Site parameters are pooled by empirical Bayes across neighboring vertices using a mesh-graph prior. New sites are handled by the method-of-moments step of CB-Predict, so a site with 3 to 20 scans can be harmonized without refitting. The radial distance scalar is harmonized jointly. This is the missing piece noted in the file: existing harmonization methods either treat scalar regional summaries, as CB-Predict does, or operate on voxel images before surfaces exist, as in 3D CycleGAN harmonization and MISPEL.
The third component is partition-marginalized inference. Final biomarker maps are reported as selection probabilities over the partition distribution, not as a single patch set. Final classifiers average predictions over partitions. This also addresses the interpretability gap the PASCS-MP authors acknowledge, because a selection probability is attached to individual vertices.
Baselines
The naive reference point is hippocampal volume plus age, harmonized with standard ComBat. The file reports volume-and-age AUCs of 0.76, 0.66, and 0.58 against surface-statistics AUCs of 0.94, 0.93, and 0.92 for the three binary comparisons in the single-site Xuanwu cohort (Classification of Alzheimer's disease based on hippocampal multivariate morphometry statistics). Our method must stay above the surface-statistics level, not merely above the volume baseline.
The strongest existing surface method for the amyloid task is PASCS-MP with its published single patch layout and no explicit harmonization. It reports amyloid-positivity accuracy of 0.89 in ADNI MCI, 0.79 in ADNI cognitively unimpaired subjects, and 0.81 in external OASIS cognitively unimpaired subjects. We will reimplement PASCS-MP on our own processed ADNI and OASIS meshes with identical train/test splits, the same random-forest classifier, and the same hyperparameter search budget. We report both its published numbers and our reproduction, and judge against the reproduction. Beating it means two things together. On external OASIS, the partition-marginalized harmonized model must exceed the reproduced PASCS-MP accuracy, with a confidence interval from repeated splits that excludes it. And the across-partition standard deviation of accuracy must shrink substantially.
For AD-versus-control detection across ADNI, AIBL, and OASIS, the strongest multi-center method in the file is UG-FedDA. It is a transformer over fixed region-of-interest features with uncertainty-weighted cross-site alignment. It reports 90.54% (ADNI), 89.04% (AIBL), and 77.78% (OASIS) normal-versus-AD accuracy, with OASIS sensitivity of 78.03% (UG-FedDA). Its features come from whole-brain regions while ours are hippocampal only. We therefore treat it as a reference ceiling rather than a like-for-like competitor. We will also run a variant that feeds our harmonized partition-marginalized features into its uncertainty-weighted classifier.
Three ablations isolate the key ingredient:
- Aggregate-then-harmonize: CB-Predict applied to patch summaries under each partition. This is the direct test of the mechanism.
- Single-partition harmonized model: removes the marginalization over partitions.
- Vertex-level harmonization without the mesh-graph prior: tests whether spatial pooling of site parameters matters at small site sizes.
Every variant uses the same subjects, folds, and tuning budget.
Why not just…?
Why not just report a sensitivity sweep over a few patch sizes? A sweep would reveal fragility but would not explain or remove it. Our mechanism predicts something a sweep cannot: the spread should concentrate where site scale factors vary within patches, and it should vanish under vertex-level harmonization. The audit component includes the sweep. The claim stands or falls on whether harmonize-before-aggregate shrinks the spread more than aggregate-then-harmonize does.
Why not just run per-vertex ComBat on log tensors and skip partitions entirely? This is the most serious threat, because it is the core of our second component. Per-vertex ComBat alone may give most of the transfer gain. Downstream models still need dimension reduction, however, and any patch or sparse-coding step reintroduces a partition choice. The ablation "vertex harmonization, single partition" against "vertex harmonization, marginalized" settles whether marginalization adds anything. If it adds nothing to accuracy, we will report that and keep marginalization only for the stability of the biomarker maps.
Why not just fit a spatially smoothed vertex-level model, such as a graph-Laplacian penalty, instead of patches? The cFSGL progression model in Feature selective temporal prediction of Alzheimer's disease progression explicitly calls for this. It reduced 48-month ADAS-Cog error from 4.367 to 3.164 while still ignoring mesh adjacency. Smoothing replaces patch size with a bandwidth, which is the scale effect under another name. We include a Laplacian-penalized vertex model and sweep its bandwidth. If its conclusions are bandwidth-stable and it matches our accuracy, it wins, and the paper becomes a recommendation to abandon patches.
Why not harmonize the raw MRI with a CycleGAN before surface extraction? Image-level harmonization requires pooling raw scans, which privacy rules often forbid, as UG-FedDA notes. It also gives no guarantee that the deformation tensors computed afterwards are site-free, since registration errors arise after harmonization. We include it only as a secondary check on the ADNI scanner subsets, where pooling is permitted.
Proposed Experiments and Decisive Checks
The data are hippocampal surfaces from ADNI, used with its real scanner and site labels rather than the synthetic institution splits of FMFS, plus OASIS and AIBL. Meshes come from FSL FIRST segmentation followed by conformal slit mapping and fluid registration, as in the prior pipelines. There are two tasks: amyloid positivity (ADNI to OASIS) and NC/MCI/AD classification (leave-one-cohort-out across all three cohorts). We also use leave-one-site-out within ADNI with target sites reduced to 3 to 20 subjects, mirroring the CB-Predict evaluation.
There are four metrics:
- External accuracy and AUC.
- Across-partition standard deviation of accuracy.
- Vertex-wise selection-probability entropy.
- Dice overlap of the selected regions between cohorts.
The first decisive check is a simulation. We plant an anisotropic deformation with zero change in local area, which no Jacobian-based or volume-based measure can see, inside CA1. We then add a spatially varying site scale. Aggregate-then-harmonize should recover the planted region with partition-dependent accuracy. Harmonize-then-aggregate should recover it uniformly, with the gap growing as predicted by the variance expression above.
The second check is on real data. Partition spread of the CA1/subiculum finding should correlate with the estimated within-patch variability of $\delta_{v,s}$.
The third check is the external OASIS amyloid comparison against reproduced PASCS-MP. If CA1 and the subiculum keep high selection probability under every partition after harmonization, the field's central localization is vindicated. If they do not, a widely cited result is partly an artifact of the patch grid. Either outcome is worth knowing.
References
- Multivariate Statistics of Tensor-Based Cortical Surface Morphometry
- Riemannian Nonlinear Mixed Effects Models: Analyzing Longitudinal Deformations in Neuroimaging
- ComBat-Predict enhances generalizability of neuroimaging models to new sites
- Fairness by "Where": A Statistically-Robust and Model-Agnostic Bi-level Learning Framework
- Classification of Alzheimer's disease based on hippocampal multivariate morphometry statistics
- Predicting Brain Amyloid Using Multivariate Morphometry Statistics, Sparse Coding, and Correntropy: Validation in 1,101 Individuals From the ADNI and OASIS Databases
- Federated Morphometry Feature Selection for Hippocampal Morphometry Associated Beta-Amyloid and Tau Pathology
- UNSUPERVISED HARMONIZATION OF BRAIN MRI USING 3D CYCLE GANS AND ITS EFFECT ON BRAIN AGE PREDICTION
- MISPEL: A supervised deep learning harmonization method for multi-scanner neuroimaging data
- UG-FedDA: Uncertainty-Guided Federated Domain Adaptation for Multi-Center Alzheimer's Disease Detection
- Feature selective temporal prediction of Alzheimer's disease progression using hippocampus surface morphometry