Brain-imaging studies increasingly pool scans from many hospitals, and each scanner leaves its own fingerprint on the measurements. Separately, the field has learned that cortical volume is misleading: it is roughly the product of cortical thickness and surface area, two quantities with different developmental origins, so opposite changes in them can cancel and hide real effects. The usual fix for scanner differences, ComBat (a statistical method that subtracts a per-site mean shift and rescales per-site variance for each brain feature), was built for the averaged view. It assumes one scanner affects every person's measurement of a region in the same way. That assumption is physically doubtful for thickness. Thickness is the distance between two boundaries the software places by reading gray–white tissue contrast, and that contrast varies across people with age, myelination and disease. A scanner's bias on thickness should therefore depend on each person's local tissue contrast. Surface area is driven more by geometric distortion and should be less sensitive to it.
We propose to test this claim and correct for it. The prediction is that scanner effects are mostly carried by thickness rather than area, and that they grow or shrink with measurable local gray–white contrast. If so, harmonizing volume, or harmonizing thickness and area with one fixed shift per site, leaves residual scanner signal that varies with tissue state. That residue can create false thickness–area dissociations, or hide real ones, especially when scanner and diagnosis are confounded. The fix we propose is simple on purpose: harmonize thickness and area separately, and let the site effect on thickness vary with a per-vertex contrast measure. The paper's contribution is to show whether this matters, by how much, and where.
Core Claim
The idea borrows a lesson from single-cell biology. There, averaging over a tissue hides cell-type-specific signal (Genome 2011; COVID-19 tissue atlases), and batch correction that ignores that heterogeneity either leaves residue or erases biology (Trade-off between conservation of biological variation and batch effect removal). We claim the same thing happens twice in multi-site morphometry. Volume averages over two geometric channels. Per-site ComBat then averages the scanner effect over people whose tissue contrast differs.
Concretely, write $T$ for thickness, $A$ for area and $V$ for volume at a vertex or region. Then $\log V \approx \log T + \log A$, so any harmonization of $V$ implicitly assigns the scanner effect to some mix of the two channels. We model thickness for subject $i$, site $s$ and region $v$ as
$$\log T_{iv} = \alpha_v + x_i^\top \beta_v + \gamma_{sv} + \delta_{sv}\, c_{iv} + \sigma_{sv}\,\varepsilon_{iv}.$$
Here $x_i$ holds the biological covariates to be protected (age, sex, diagnosis). $c_{iv}$ is the subject's gray–white contrast ratio sampled at the white-matter boundary. $\gamma_{sv}$ and $\sigma_{sv}$ are the usual ComBat site location and scale terms. $\delta_{sv}$ is the new site-by-contrast term. Area gets the same model, and we expect its $\delta$ to be near zero. Standard ComBat is the special case $\delta=0$. The empirical question is whether $\delta_{sv}$ is reliably non-zero and whether removing it changes scientific conclusions.
Relation to Prior Work
The morphometry literature motivates the split into channels. Separate analysis of thickness and area outperforms volume (Joint Analysis of Cortical Area and Thickness). The two channels dissociate clinically in conduct disorder (Cortical thickness, surface area, and folding alterations) and in sex differences in depression (Sex differences of brain cortical structure in MDD). They also change in non-overlapping zones over time (Cross-sectional and longitudinal neuroanatomical profiles in autism). None of these studies asks whether the reported dissociations survive channel-specific scanner effects, and the research file notes this gap.
The harmonization literature supplies the method and the main competitor. Cluster ComBat shows that site parameters fall into a few clusters, which lets it generalize to unseen sites on 563 ADNI subjects from 18 sites. It still applies one linear shift and scale per cluster to pre-extracted features. We keep its clustering idea for unseen sites and add the contrast-dependent term.
Deep invariance methods are the more flexible alternative. These include adversarial batch-free autoencoders (Batch Effect Removal via Batch-Free Encoding), conditional-prior latent splitting (inVAE), and gradient-reversal training (Generalized and Invariant Single-Neuron In-Vivo Activity Representation Learning). The inVAE paper and the trade-off paper also act as constraints on our design. They show that invariance objectives erase biology when batch and biology are correlated. We therefore keep an explicit, inspectable linear model with protected covariates rather than an adversarial latent code.
Imaging-transcriptomics work defines the downstream use case. The MDD subtype study (Cortical Morphometric Inverse Divergence) combines thickness, area and other features into one similarity network. It applies standard ComBat on the multi-site SRPBS collection, then regresses the result on bulk Allen Human Brain Atlas expression. Our proposal asks whether its gene associations look different when the network is rebuilt per channel after contrast-aware harmonization.
Research Program
The work has three stages.
First, we measure the scanner effect. We run one fixed surface-reconstruction pipeline on ADNI, SRPBS and, where access permits, the ENIGMA schizophrenia cohort (Obesity and brain structure in schizophrenia). For each scan we extract per-region thickness, area, volume and gray–white contrast. We then estimate $\gamma$, $\delta$ and $\sigma$ per channel. The key descriptive results are the proportion of total site variance carried by thickness versus area, and the fraction of regions where $\delta_{sv}$ is significant after correction for multiple comparisons.
The strongest evidence comes from subjects scanned on more than one scanner, for example ADNI participants imaged across scanner changes and the SRPBS traveling-subject scans if access permits. For these people, true anatomy is fixed. Any between-scanner difference is pure artifact, and we can test directly whether it tracks contrast.
Second, we build the corrector. One version is the linear model above, fitted with ComBat-style empirical Bayes pooling across regions. The other replaces $\delta_{sv} c_{iv}$ with a smooth function of contrast fitted by penalized splines. Both versions inherit Cluster ComBat's assignment of new sites to parameter clusters, so they generalize to sites never seen in training.
Third, we test whether the correction changes conclusions. We recompute the MDD subtype analysis and the thickness-versus-area dissociations under each harmonization, and report which findings survive.
Baselines
Raw features. The naive reference is thickness, area and volume with no harmonization. It anchors how much site signal exists at all.
Volume-level ComBat. This is standard ComBat applied to regional volume, the common practice. Thickness and area are then derived by assuming the correction splits equally between the two channels in log space. It represents the averaged view we argue against.
Per-channel ComBat. Standard ComBat applied separately to thickness and area, with age, sex and diagnosis protected. This is the simplest strong competitor. It shares our channel split and lacks only the contrast term, so comparing it with our method is the ablation that isolates the key ingredient.
Per-channel ComBat with contrast as an ordinary covariate. Contrast enters as a protected covariate with one slope shared across sites, but there is no site-by-contrast interaction. This isolates whether the site dependence of the slope matters, rather than contrast itself.
Cluster ComBat. Cluster ComBat is the strongest published method for this data type and the only one in the file evaluated on ADNI multi-site morphometry with unseen-site generalization. The paper reports lower downstream mean absolute error than ComBat and reduced site classification accuracy. No reported numeric MAE or accuracy value was found on Lacuna for Cluster ComBat on ADNI; the only figures given are 563 subjects, 18 sites, 2x and 4x speedups, and stability across $k \in \{3,5,7,9\}$. We therefore rerun it ourselves on identical feature tables and splits, sweeping the cluster count $k$ over the same range.
Adversarial autoencoder. An adversarial autoencoder in the style of Batch-Free Encoding is trained on the concatenated thickness and area vectors. It represents flexible nonlinear correction and gets the same hyperparameter search budget as the spline version of our method.
Fairness and the bar to clear. All methods see the same subjects, the same protected covariates and the same leave-sites-out splits. Hyperparameters are tuned only on training sites. Our method beats the field only if it meets three conditions against per-channel ComBat and Cluster ComBat. It must reduce cross-scanner disagreement in same-subject scans, with a confidence interval excluding zero. It must reduce held-out site classification accuracy. And it must do both while keeping the recovery of known age effects and simulated diagnosis effects within noise of the uncorrected effect sizes.
Why not just…?
Why not just run ComBat separately on thickness and area? This is the real threat. If $\delta_{sv}$ turns out to be negligible, per-channel ComBat is enough, and our contribution shrinks to a careful demonstration that channel splitting matters. That would still be useful, but it is smaller than what we claim. The same-subject cross-scanner analysis settles the question directly: if the residual cross-scanner thickness differences left after per-channel ComBat do not correlate with contrast, we will report that and drop the interaction term.
Why not just add contrast as a covariate in ComBat? A shared slope treats contrast as biology and protects it, so it cannot remove a bias that differs between scanners. Worse, contrast changes with age and disease, so protecting it with a shared slope may lock scanner artifact into the "biological" part. The covariate-only baseline tests this directly. If it matches the interaction model on cross-scanner agreement, the site dependence of the slope is unnecessary.
Why not just use Cluster ComBat or a deep harmonizer? Cluster ComBat solves a different problem: how to share parameters across sites. It still applies one shift per cluster and so cannot follow differences in tissue state within a site. We include its clustering inside our method rather than competing with it. Deep adversarial correction can in principle learn the contrast dependence. However, the trade-off study and inVAE document that such correction erases biology when site and biology are confounded. Our confounded-simulation experiment measures exactly this failure.
Proposed Experiments
The first decisive check is a null-dissociation simulation. Starting from real ADNI features, we build cohorts with no true diagnosis effect, then assign "patients" mostly to some scanners. We count how often each harmonization yields a false thickness-only or area-only group difference. The prediction is that volume-level ComBat and per-channel ComBat produce excess false thickness findings, concentrated in high-myelin regions where contrast differs most, and that the contrast-aware model does not.
The second check is an injected-signal simulation. We plant a known thinning paired with an area expansion that leaves volume unchanged, mimicking the cancellation described in Joint Analysis of Cortical Area and Thickness. We then measure how well each method recovers both components under site confounding.
The third check uses the same-subject scans described above, which give direct ground truth on artifact: for these people, every residual cross-scanner difference after harmonization is error.
Finally, as the applied test, we repeat the SRPBS MDD subtyping from the MIND study on 240 patients and 367 controls, with similarity networks built separately for thickness and area. We report whether the two-subtype structure (78 and 162 patients) and the gene associations persist. A finding that the reported subtypes are partly driven by thickness and partly by scanner-contrast residue would be the clearest evidence that harmonization hides heterogeneity in the same way bulk averaging does.
References
- As we come to the end of 2011, several members of the Genome Biology Editorial Board give their views on the state of play in genomics
- COVID-19 tissue atlases reveal SARS-CoV-2 pathology and cellular targets
- Trade-off between conservation of biological variation and batch effect removal in deep generative modeling for single-cell transcriptomics
- Joint Analysis of Cortical Area and Thickness as a Replacement for the Analysis of the Volume of the Cerebral Cortex
- Cortical thickness, surface area, and folding alterations in male youths with conduct disorder and varying levels of callous–unemotional traits
- Sex differences of brain cortical structure in major depressive disorder
- Cross-sectional and longitudinal neuroanatomical profiles of distinct clinical (adaptive) outcomes in autism
- Distributed Harmonization: Federated Clustered Batch Effect Adjustment and Generalization
- Batch Effect Removal via Batch-Free Encoding
- Conditionally Invariant Representation Learning for Disentangling Cellular Heterogeneity
- Generalized and Invariant Single-Neuron In-Vivo Activity Representation Learning
- Molecular Mechanisms Explaining Neuroanatomical Subtypes in Major Depressive Disorder: Insights From Cortical Morphometric Inverse Divergence
- Obesity and brain structure in schizophrenia – ENIGMA study in 3021 individuals