Robust brain age estimation from structural MRI with contrastive learning Journal: Pattern Recognition Letters
Carlo Alberto Barbano, Benoit Dufumier, Edouard Duchesnay, Marco Grangetto, Pietro Gori, for the Alzheimer’s Disease Neuroimaging Initiative · 2025
Read Robust brain age estimation from structural MRI with contrastive learning Journal: Pattern Recognition Letters by Carlo Alberto Barbano, Benoit Dufumier, Edouard Duchesnay, Marco Grangetto, Pietro Gori, for the Alzheimer’s Disease Neuroimaging Initiative (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2507-01794.txt.
Introduction
Carlo Alberto Barbano and four co-authors estimate the brain's age from MRI scans using contrastive learning, in a 2025 paper. The catalog title carries stray journal text; the printed title is clean. A brain that looks older than its years signals risk for decline.
Instead of training directly on age labels, their model learns by comparing scans, with a new loss function built for the job. Across more than 20,000 scans from many sites, wider pre-training nearly halved the error on outside data. The loss resisted scanner-related quirks. It also captured faster aging in cognitive impairment and Alzheimer's disease. Accuracy tracked diagnostic usefulness, unlike label-trained rivals. The authors pitch it as a foundation for neuroimaging models.
Catalog record
| CREATOR | Carlo Alberto Barbano, Benoit Dufumier, Edouard Duchesnay, Marco Grangetto, Pietro Gori, for the Alzheimer’s Disease Neuroimaging Initiative |
|---|---|
| YEAR | 2025 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 30 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2507-01794
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
Robust brain age estimation from structural MRI with contrastive learning Journal: Pattern Recognition Letters Carlo Alberto Barbano, Benoit Dufumier, Edouard Duchesnay, Marco Grangetto, Pietro Gori, for the Alzheimer’s Disease Neuroimaging Initiative 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 supervised approaches for brain age estimation. We introduce a novel contrastive loss function, ℒ e x p \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, ℒ e x p \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 supervised baselines, ℒ e x p \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. 1 Introduction Brain aging is a complex and individualized process, involving structural changes such as cortical thinning and white matter loss, which vary widely across the population. Accurately modeling these subject-specific trajectories is a long-standing challenge in neuroscience, with high clinical relevance for detecting and tracking age-related neurodegenerative diseases. Recent deep learning (DL) approaches have shown promise in building brain-age predictors that learn mappings from anatomical MRI to chronological age ( Peng et al. 2021 ) . However, these models typically require large-scale training data, which necessitates the aggregation of scans from multiple imaging sites and studies. A major challenge with multi-site data is that DL models, especially deep neural networks, tend to overfit to site-specific artifacts, such as differences in scanner hardware, acquisition protocols, or magnetic field strength. This phenomenon leads to degraded performance on unseen sites and severely limits real-world applicability ( Glocker et al. 2019 ; Wachinger et al. 2021 ) . As a result, the development of models that generalize well across heterogeneous sources remains an open and pressing problem. The OpenBHB challenge ( Dufumier et al. 2022 ) was introduced as a benchmark to address this issue, offering a large-scale, multi-site dataset designed to evaluate brain-age prediction under real-world generalization constraints. Traditional models participating in this benchmark have largely relied on supervised regression objectives, such as mean absolute error (MAE) ( Cole et al. 2017 ; Jónsson et al. 2019 ; Peng et al. 2021 ) . However, these frameworks typically ignore the presence of site-specific confounds during training, which may affect the learned representations. To address this, we previously introduced a Contrastive Learning (CL) framework for brain-age regression ( Barbano et al. 2023a ) , based on a novel similarity-driven loss function, \mathcal{L}^{exp} , which encourages biologically meaningful and site-invariant embeddings. In this extended work, we scale up our previous approach both in data and scope: we employ over 20,000 structural T1 scans from four large public datasets (OpenBHB, ADNI, ICBM, and OASIS-3), and conduct a broad empirical study to analyze generalization, site robustness, clinical validity, and downstream transferability. Our contributions are as follows: 1. We significantly improve the state-of-the-art performance in brain age prediction on the OpenBHB challenge; 2. We show that our proposed formulation \mathcal{L}^{exp} achieves consistent invariance towards site bias; 3. We show that CL is a reliable estimator of accelerated brain aging in patients with cognitive impairment (MCI) and Alzheimer’s disease (AD); 4. We show that better brain age estimators based on CL can result in better diagnostic models. 2 Results Table 1 : Brain age regression results. Single means that the models were trained and tested on each dataset separately; Merge means that the models were trained on the combined healthy samples. OpenBHB Challenge Results report balanced accuracy (BAcc), internal MAE, external MAE, and the composite score \mathcal{L}_{c} . OpenBHB ADNI ICBM OASIS3 OpenBHB Challenge Results Method Train MAE R^{2} MAE R^{2} MAE R^{2} MAE R^{2} BAcc Int. MAE Ext. MAE \mathcal{L}_{c}\,(\downarrow) L1 Single 2.73 0.92 3.64 0.58 4.33 0.75 4.32 0.65 6.7 2.67 4.18 1.82 \mathcal{L}^{y-aware} ( Dufumier et al. 2021 ) 2.65 0.92 3.47 0.61 4.21 0.76 4.26 0.65 6.60 2.66 4.10 1.82 \mathcal{L}^{threshold} ( Barbano et al. 2023a ) 2.66 0.93 3.36 0.63 5.27 0.57 4.18 0.66 5.73 2.95 4.10 1.74 \mathcal{L}^{exp} ( Barbano et al. 2023a ) 2.66 0.93 3.44 0.62 4.11 0.77 4.22 0.66 5.1 2.55 3.76 1.54 L1 Merge 2.59 0.93 3.45 0.61 3.77 0.82 4.08 0.68 9.33 2.42 2.43 1.19 \mathcal{L}^{y-aware} ( Dufumier et al. 2021 ) 2.70 0.91 3.40 0.63 3.73 0.82 4.08 0.67 7.3 2.61 2.53 1.15 \mathcal{L}^{threshold} ( Barbano et al. 2023a ) 3.11 0.85 3.43 0.62 3.83 0.80 4.11 0.66 6.0 2.90 2.21 0.95 \mathcal{L}^{exp} ( Barbano et al. 2023a ) 2.77 0.91 3.39 0.63 3.84 0.81 4.05 0.68 5.40 2.71 2.25 0.93 (a) External MAE vs training size (scaling) (b) Site bias vs training size. \mathcal{L}^{exp} stays consistent. Figure 1 : Scaling contrastive brain-age models with data size. (a) External mean absolute error (MAE) decreases as training size increases; all methods 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. (a) Age delta (BAG) distribution for HC, MCI, and AD classes across all datasets. The shaded areas represent the age delta distribution (histogram) for each class. The horizontal black line in the middle highlights the mean value (explicitly annotated for readability). (b) ROC plots for using BAG as a proxy for detecting Alzheimer’s disease (AD) patients. (c) Longitudinal change in BAG on ADNI. We only considered patients with a minimum of 3 follow-up exams. Figure 2 : Using brain age gap (BAG) as a proxy for detecting neurodegeneration. (a) Age delta (BAG) distribution for healthy controls (HC), stable MCI (sMCI), progressive MCI (pMCI), and Alzheimer’s disease (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. We base our analysis on a collection of different publicly available datasets (OpenBHB, ADNI, ICBM, OASIS) for a total of 7,045 subjects and 20,647 structural MRI scans. 2.1 Scaling training size improves generalization In Tab. 1 we report results of brain age regression across the different datasets employed. We experiment with two pre-training schemes, namely, we first train and test on the same dataset (single), and then we