Synthetic Brain Scans Bridged Between Sites Sharpen Brain Age Prediction

Magnetic resonance imaging has become one of the most powerful windows into how the human brain ages, but it carries a stubborn flaw: every scanner tells the story slightly differently. Even two machines of the same model from the same manufacturer produce systematic differences in image intensities, tissue volume measures and connectivity, driven by variations in scanner type, imaging protocol, magnetic field strength, acquisition parameters and reconstruction algorithms. Researchers call this the “Effects of Site,” and for brain age prediction—a machine learning technique that estimates a person’s biological brain age from an MRI scan and compares it with chronological age—these scanner-specific biases can drown out the very biological signals scientists are trying to detect. A new open-access study, published on 10 March 2026 in the journal Discover Data, proposes a strikingly simple fix: instead of statistically scrubbing site effects out of the data, generate new synthetic training samples by interpolating between demographically matched participants scanned at different sites.

The method, dubbed Inter-Site SMOTE, is an adaptation of the synthetic minority over-sampling technique, a classic data augmentation tool. But where standard SMOTE interpolates between neighboring samples within the same class, the new approach draws its synthetic samples from pairs of participants at two different imaging sites who are strictly matched by age and gender. The logic is elegant: matched pairs should share similar biological signals, because their brains are being compared across people of the same sex and nearly the same age, while their site-specific noise profiles differ. Averaging between them, the authors hypothesized, preserves the shared biology while cancelling out scanner artifacts—effectively manufacturing training examples that live in the domain-invariant space between scanners.

The mathematics is straightforward. Given a feature vector from a base site and one from a target site, a synthetic sample is created by linear interpolation, with a weighting factor drawn uniformly between zero and 0.5. Restricting the weight to this range keeps every synthetic sample in the local neighborhood of the base sample, so the procedure acts as a kind of “site-regularized” augmentation rather than a wild extrapolation. The synthetic sample inherits the base site’s label, the shared gender, and an age computed with the same interpolation weight, keeping the metadata consistent with the features. The feature vectors themselves were modulated gray matter volume maps extracted with voxel-based morphometry and resampled to eight-millimeter voxels, yielding 3,747 features per image.

To test whether this interpolation really strips out scanner signatures while retaining biological information, the team first trained two models on pooled data from four sites: a linear support vector classifier to predict which site a scan came from, and a kernel ridge regressor to predict age. They then generated synthetic test sets at a series of interpolation weights and evaluated both models on them. Meanwhile, the mean absolute error of age prediction on the synthetic data decreased with increasing interpolation weight, showing that the age-related signal was not erased but actually better retained as the samples moved toward the “site-neutral” zone between scanners.

The decisive test came in a leave-one-site-out scheme, the gold standard for measuring generalization to unseen scanners. Data from three of four sites trained a kernel ridge regressor, which was then evaluated on the held-out site, with each site taking a turn as the test set. Three training scenarios were compared: a baseline pooling only the original data, a model trained exclusively on synthetic data, and a hybrid model trained on original plus synthetic data, alongside a comparison against ComBat, the most widely used statistical harmonization method. The hybrid approach delivered the best results, producing a significant reduction in mean absolute error for the test sites IXI/HH and IXI/Guys and comparable performance for eNKI and CamCAN. Models trained purely on synthetic data started with high errors but improved drastically as more synthetic samples were added, suggesting that large volumes of synthetic data are needed to cover the variance of the original observations.

The head-to-head with ComBat was revealing. In the IXI/HH and IXI/Guys test sets, ComBat harmonization actually increased the mean absolute error to roughly 6.18 and 6.4 years respectively, compared with about 5.55 and 5.65 for the unharmonized baseline, with a smaller degradation also seen for CamCAN. Only for the eNKI site did ComBat help, reaching an error of about 6.4 years that beat the baseline of 6.55 and matched the proposed method. The authors attribute this failure to a fundamental conflict between statistical harmonization and predictive modeling: ComBat needs biological covariates—in this case age—modeled during harmonization to preserve biological variance, but in a prediction task age is the unknown target. Using it to harmonize the test set would constitute data leakage, while omitting it risks removing the very signal of interest. Statistical harmonization is also sensitive to sample size imbalances and cannot be applied to data from an unseen site at all.

To rule out the possibility that the gains were merely a side effect of having more training data, the researchers designed a clever control experiment. They split the largest site, eNKI, into two age-matched subsets and used CamCAN as an additional matched site. Synthetic data was then generated in two ways: by interpolating within the same site, between the two eNKI subsets, or across sites, between eNKI and CamCAN. Both augmented datasets were tested on the three unseen IXI sites, with the whole process repeated twenty times to account for randomness. Across all upsampling levels, cross-site synthetic data led to better mean absolute error on average than within-site synthetic data, with the advantage clearest at the highest upsampling level for IXI/Guys and IXI/IOP and negligible for IXI/HH. Within-site augmentation provided only a marginal benefit by densifying the training distribution; it failed to remove site effects. The improvement, the authors conclude, comes from the specific geometry of cross-site interpolation—bridging the domain gap—rather than simple regularization through extra samples.

The datasets underpinning the study were three large neuroimaging collections of healthy adults spanning ages 18 to 88: the Cambridge Center for Ageing and Neuroscience (CamCAN), the Information eXtraction from Images (IXI) dataset, and the enhanced Nathan Kline Institute-Rockland sample (eNKI), together comprising 2,031 participants. Because the IXI data were acquired at three different hospitals on different scanners, each center was treated as a separate site, giving four sites in total. The smallest of these, IXI-IOP, was used only as a test site and never for training. The study’s scope is important: all participants were healthy adults, the imaging modality was T1-weighted MRI reduced to gray matter volume features, and the prediction model was a kernel ridge regressor, though the authors note the harmonization framework itself is model-agnostic.

The method is not without limitations, which the authors discuss candidly. Linear interpolation assumes the data manifold is locally linear within small neighborhoods, but brain aging trajectories and scanner-induced variations likely live on complex, non-linear manifolds in high-dimensional feature space. There is a theoretical risk that linearly interpolated samples represent biologically implausible combinations of regional brain volumes, and the observed gains may partly reflect the synthetic data acting as a form of site-aware regularization that smooths the decision boundary across the domain gap, rather than perfectly reconstructing biological anatomy. Matching strictly on age and gender also imposes constraints: as more covariates—such as ethnicity, education level or clinical status—are added to the matching process, finding exact matches across sites becomes progressively harder, demanding larger datasets. The increased sample size can also strain computationally expensive kernel methods, whose cost grows cubically with data volume.

Even with those caveats, the study offers a pragmatic path forward for a field whose clinical ambitions increasingly depend on multi-center data. Brain age gap estimates have shown promise as biomarkers for detecting accelerated brain aging, with associations reported with cognitive impairment and Alzheimer’s disease, and potential utility in early detection of neurodegenerative conditions such as Alzheimer’s and Parkinson’s disease. But deploying such models in practice requires them to work on scanners they have never seen. By generating synthetic samples that populate the sparse regions between site distributions, Inter-Site SMOTE improves cross-scanner generalization without the assumptions, leakage risks and unseen-site limitations of statistical adjustment methods, and without the enormous data appetite of deep learning harmonization approaches. The authors have released their Python code publicly, and future work will explore richer covariate matching, application to deep learning architectures, other neuroimaging modalities, and sensitivity to the method's hyperparameters. For now, the message is a compelling one for the age of big-science brain imaging: sometimes the best way to harmonize your data is not to subtract the noise, but to average your way straight through it.

Subject of Research: A new data harmonization method that generates synthetic MRI training samples by interpolating between matched participants from different scanning sites to improve brain age prediction across unseen scanners.

Article Title: Data harmonizing via interpolation applied to brain age prediction

Article References: Nieto, N., Asati, A., Jadhav, K., & Patil, K. R. (2026). Data harmonizing via interpolation applied to brain age prediction. Discover Data, 4(1), Article 3. https://doi.org/10.1007/s44248-026-00100-7

Image Credits: AI Generated

DOI: 10.1007/s44248-026-00100-7

Keywords: brain age prediction, MRI, data harmonization, Inter-Site SMOTE, site effects, machine learning, ComBat, neuroimaging, synthetic data, leave-one-site-out, gray matter volume, generalization