Robust brain age estimation from structural MRI with contrastive learning

TL;DR

The paper tackles cross-site variability in brain-age estimation from structural MRI by applying contrastive learning with a novel loss, $L^{exp}$, to learn site-invariant embeddings. It scales pretraining to over 20,000 scans across OpenBHB, ADNI, ICBM, and OASIS-3 and demonstrates improved external MAE, reduced site bias, and robust detection of accelerated aging in neurodegenerative conditions. The study shows that contrastive pretraining enhances transfer to downstream diagnostic tasks and maintains meaningful clinical signals, suggesting contrastive learning as a promising foundation for neuroimaging models. Collectively, the work provides evidence that large, heterogeneous pretraining data paired with the $L^{exp}$ loss yield robust, generalizable brain representations with real clinical impact.

Abstract

Estimating brain age from structural MRI has emerged as a powerful tool for characterizing normative and pathological aging. In this work, we explore contrastive learning as a scalable and robust alternative to L1-supervised approaches for brain age estimation. We introduce a novel contrastive loss function, $\mathcal{L}^{exp}$, and evaluate it across multiple public neuroimaging datasets comprising over 20,000 scans. Our experiments reveal four key findings. First, scaling pre-training on diverse, multi-site data consistently improves generalization performance, cutting external mean absolute error (MAE) nearly in half. Second, $\mathcal{L}^{exp}$ is robust to site-related confounds, maintaining low scanner-predictability as training size increases. Third, contrastive models reliably capture accelerated aging in patients with cognitive impairment and Alzheimer's disease, as shown through brain age gap analysis, ROC curves, and longitudinal trends. Lastly, unlike L1-supervised baselines, $\mathcal{L}^{exp}$ maintains a strong correlation between brain age accuracy and downstream diagnostic performance, supporting its potential as a foundation model for neuroimaging. These results position contrastive learning as a promising direction for building generalizable and clinically meaningful brain representations.

Figures (4)

• Figure 1: Scaling contrastive brain-age models with data size.(a) External mean absolute error (MAE) decreases as training size increases; all methods besides RnC benefit from more data, reaching a MAE between 2.21 and 2.53. (b) Balanced accuracy (B. Acc) in predicting scanner site increases for most methods as training size grows, indicating increased site bias. In contrast, $\mathcal{L}^{exp}$ remains stable and low across all sizes, demonstrating robustness to site-specific confounds. On the other hand, the decrease in B. Acc for RnC can be attributed to the degraded performance in brain age prediction when datasets are merged (e.g., the model collapsed and did not learn meaningful representations). In fact RnC shows a trend similar to other methods when trained on OpenBHB only. We hypothesize this is due to the unweighted alignment term and to the high temperature value used by RnC, which could focus the repulsion on samples coming from different dstributions (e.g., ADNI and OASIS subjects are older), leading to a latent space which does not capture well biological and age variability.
• Figure 2: Using BAG as a proxy for detecting neurodegeneration. (a) BAG distribution for HC, stable MCI (sMCI), progressive MCI (pMCI), and AD patients across all datasets. All methods consistently show increased BAG in pMCI and AD groups, indicating accelerated brain aging. (b) ROC curves evaluating the ability of BAG to separate HC and AD subjects in ADNI and OASIS datasets. Contrastive methods ($\mathcal{L}^{exp}, \mathcal{L}^{threshold}, \mathcal{L}^{yaware}$) match or slightly outperform the L1 baseline. (c) Longitudinal BAG trajectories (with $\mathcal{L}^{exp})$ on ADNI, showing that pMCI and AD subjects exhibit increased BAG over time, while HC and sMCI groups remain stable. This confirms that contrastive models capture disease progression dynamics.
• Figure 3: Calibration curves show that all fine-tuned models are reasonably calibrated on the downstream task, with an Average Calibration Error (ECE) slightly higher than 10%.
• Figure 4: Correlation between downstream accuracy (ADNI) and brain age prediction error. Contrastive methods such as $\mathcal{L}^{exp}$ benefit from lower MAE.