acceptodds
Under review as a conference paper at ICLR 2027

MT-ReBA: Anatomically Grounded Regional Brain Age Representation Learning via Weak Supervision

Abstract

Brain age estimation provides a measure of brain aging, but most existing methods produce only global estimates, obscuring regional aging heterogeneity. Learning regional aging representations remains challenging because chronological age is defined at the subject level, while regional biological ages are neither directly observable nor supported by regional annotations. We introduce Multi-Task Regional Brain Age Estimation (MT-ReBA), a weakly supervised framework for learning anatomically grounded regional aging representations from T1-weighted MRI without regional labels. The key idea is to replace one atlas-defined parcel with its counterpart from another healthy subject, thereby coupling a localized image modification with a donor-age weak target and converting subject-level age labels into localized weak supervision. Specifically, an atlas-defined donor parcel is integrated into a recipient brain, and the donor's chronological age supervises the corresponding regional head as a weak proxy for regional aging. We evaluate MT-ReBA across six public cohorts. MT-ReBA achieves a global brain MAE of 3.10 years and a regional age-ranking accuracy of 69.63 ± 0.11%, shows agreement with constructed weak targets on mixed images, and exhibits anatomically specific responses and exploratory disease associations on unspliced scans.

Primary area:
applications to neuroscience & cognitive science
Keywords:
Brain Age Estimation, Weak Supervision, Neuroimaging

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_v7LYOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper tackles regional brain age estimation using a weakly supervised framework that integrates an innovative atlas-guided region mixing strategy to supervise regional heads during learning. It demonstrates that the resulting global-regional framework improves global brain age prediction and improves more localized metrics such as regional age ranking accuracy. It also shows qualitative regional BAG results on AD and PD datasets.

Critical strengths
  1. Relevant and challenging problem. The paper addresses regional brain-age estimation, an important problem given evidence that brain aging involves multiple spatially heterogeneous patterns rather than a single uniform process (Smith et al., 2020).

  2. Interesting weak-supervision strategy. The authors propose atlas-guided regional replacement to transform subject-level chronological-age labels into localized supervision for the corresponding regional prediction head. Restricting mixed-image gradients to the regional pathway is a reasonable design choice to limit the influence of synthetic-image artifacts on global age prediction. The approach also explicitly addresses the challenge of assigning anatomically meaningful outputs when regional age labels are unavailable.

  3. Multiple complementary evaluation strategies. Beyond evaluating agreement with the synthetic weak targets, the authors examine anatomical specificity on unspliced MRI using ROI occlusion and input-gradient analyses. These experiments provide useful evidence that the regional heads respond preferentially to their corresponding anatomical parcels rather than relying exclusively on the artificial replacement procedure.

Critical weaknesses
  • The problem of regional brain age estimation has been tackled in the neuroimaging literature but always suffer from the same issue: there is no proper ground-truth and the concept of regional biological age is ill-defined. While the authors propose a synthetic way to mix brain regions across subjects in order to artificially introduce brain regional age diversity, it is impossible to assume that brain regions chronological age equal regional biological age. As such, the evaluation metrics are always biased and do not measure whether the model is actually learning this "biological regional brain age". There are only weak proxy of such measures (i.e. the proposed regional age ranking accuracy).

  • The method is strongly dependent of the anatomical atlas for the mixing strategy and more experiments should be conducted to assess the sensitivity of the representation to the choice of the atlas. Since atlas-guided replacement is central to the proposed contribution, this sensitivity should be evaluated. The current restriction to 48 cortical regions also limits the scope of the proposed regional analysis, particularly given the importance of subcortical structures in neurodegenerative disorders.

  • The optimization procedure and evaluation uncertainty require further justification. The Stage 2 optimization uses separate real-image and mixed-image passes, with explicit gradient routing to prevent mixed-image losses from updating the backbone directly. While this design is understandable, the paper does not establish that separate optimization steps are necessary or superior to an implementation that combines the objectives while preserving the intended gradient restrictions. An ablation comparing these alternatives, including training stability and computational cost, would help justify the additional complexity.

  • The training dataset is modest relative to the available pretraining resources, and external generalization is not established: the pretraining dataset is rather small (<4k samples) while much larger-scale datasets are now available (i.e. OpenMind dataset with >100k samples). More importantly, the participant-level split does not establish generalization to an entirely unseen cohort. A leave-one-cohort-out evaluation or independent external validation would help determine whether the learned representations generalize across acquisition protocols and data distributions.

  • The clinical utility of the proposed regional representations is not sufficiently evaluated. The authors motivate the method partly through its potential contribution to downstream disease classification. However, the AD-versus-HC and PD-versus-HC experiments primarily report descriptive, region-wise differences in regional brain-age gaps. I would like to see quantitative classification results, including AUC and confidence intervals, for AD versus HC and PD versus HC. In particular, the authors should compare regional BAG features with a global BAG baseline, ideally also assessing whether regional features provide incremental predictive value beyond age, sex, and other relevant covariates. This would better support the paper's stated objective of demonstrating the practical utility of regional brain-age representations. Existing work on brain-structure ages for multi-disease classification (Nguyen et al., 2024) provides a relevant point of comparison.

  • Missing baselines in Table 1: direct method such as Peng et al. 2021 or ViT-based methods (e.g. Triamese-ViT, Zhang and Jiang 2024) are missing.

Questions for authors
  • What exactly is the ground truth for regional biological brain age? The method uses the donor's chronological age as a weak target for a transplanted brain region. Why should this intervention make the donor's chronological age a valid target for the region's biological age? How do the authors distinguish age-related information from inter-individual anatomical variability and other subject-specific characteristics?

  • Does the model capture genuine regional aging, or simply respond to the synthetic intervention? Can the authors demonstrate that the regional predictions reflect age-related variation in unmodified MRI independently of the region-mixing procedure? For example, do the predictions replicate in longitudinal data, track within-subject changes over time, or correlate with independent regional markers of neurodegeneration?

  • Do the regional representations provide clinically meaningful information beyond global brain age? For AD-versus-HC and PD-versus-HC, what are the AUCs and confidence intervals when using regional BAG features versus global BAG? Does adding regional features significantly improve discrimination after accounting for age and sex?

7 comments
AuthorsOpenReview ↗

Responses to Weakness 1

Thank you very much for taking the time to review and for your support. We try our best to address your questions as follows. Changes are marked in the revised PDF for easy reference.

W1

Reviewer comment. The problem of regional brain age estimation has been tackled in the neuroimaging literature but always suffer from the same issue: there is no proper ground-truth and the concept of regional biological age is ill-defined. While the authors propose a synthetic way to mix brain regions across subjects in order to artificially introduce brain regional age diversity, it is impossible to assume that brain regions chronological age equal regional biological age. As such, the evaluation metrics are always biased and do not measure whether the model is actually learning this "biological regional brain age". There are only weak proxy of such measures (i.e. the proposed regional age ranking accuracy).

Response.

Regional brain-age estimation is an emerging field, and it indeed lacks clear, unified definitions, directly observable biological ground-truth labels, and universally accepted evaluation criteria. Without regional ground-truth, establishing a reliable evaluation framework remains the primary open challenge. Our aim is to develop a systematic framework for regional brain-age learning and evaluation through weak supervision and multidimensional validation.

Crucially, as the reviewer noted, all existing works in this field face this exact limitation. For instance, Popescu et al. (2021, Frontiers in Aging Neuroscience) validated voxel-level predictions indirectly through test-retest reliability and disease-related differences. Gianchandani et al. (2024, MELBA) relied on voxel-level regional age differences between healthy and dementia patients. Fong et al. (2025, IEEE EMBC) assessed representations via hemispheric symmetry and cognitive associations. None of these studies possess direct measures or objective ground-truth to establish absolute validity; instead, they all rely on indirect, weak proxy evidence. Therefore, metrics like our RankAcc are not intended to be perfect ground-truth measures, but rather necessary proxies in the absence of absolute labels.

To address the inherent bias of any single proxy metric, we expanded and strengthened our evaluation into a rigorous multidimensional validation framework. These independent experiments converge to support that MT-ReBA learns anatomically localized, age-related representations:

1.Validity of weak supervision: Mixed-effects models on 74,700 donor comparisons yielded test-set slopes of 0.267–0.334, with zero-dose controls showing no shift (pp. 9–10, Sec 4.2.6).

2.Anatomical specificity: ROI occlusion on unspliced MRI yielded a localization accuracy of 75.0% (chance level: 2.1%, p = 0.0005) (pp. 21–22, App N).

3.Natural aging & cross-atlas robustness: In 277 independent HCs, all HO48 and AAL3 regions passed FDR correction (p. 7, Sec 4.2.3).

4.Disease relevance: Regional features significantly increased AD classification AUC (+0.128 for HO48, +0.120 for AAL3, p < 0.0005)(pp. 7–9, Secs 4.2.4–4.2.5).

AuthorsOpenReview ↗

Responses to Weakness 2

W2

Response.

To directly address your concern regarding the atlas-dependence of our atlas-guided replacement strategy and its sensitivity, we evaluated MT-ReBA using two brain atlases: Harvard–Oxford (HO48) and Automated Anatomical Labeling 3 (AAL3).

We agree that limiting the analysis to cortical regions (like the 48 regions in HO48) restricts the scope of regional analysis, especially given the critical importance of subcortical structures in neurodegenerative diseases. To resolve this, we specifically introduced AAL3, which extends to 166 regions and explicitly covers key subcortical structures such as the thalamus, substantia nigra, and locus coeruleus.

We added several experiments:

  1. If our model were overfitted to a specific atlas, its performance would degrade substantially when transferred. However, MT-ReBA maintained high regional age-ranking accuracy on both atlases (RankAcc: 68.30%–70.17%), consistently outperforming the best baseline methods by a large margin of 10.41–16.11 percentage points across three random seeds. Specifically, RankAcc remained robust at 69.63% on HO48 and 69.11% on AAL3.

These results strongly confirm that our model's regional ranking capability is driven by generalizable biological aging patterns rather than atlas-specific artificial features, demonstrating high robustness across different anatomical resolutions and cortical/subcortical boundaries.(p. 6, Section 4.2.2)

Table 2. Regional brain-age ranking accuracy (RankAcc, %). Results are mean ± SD across three training seeds.

HO48

\hline \text{Method} & \text{Seed 1} & \text{Seed 2} & \text{Seed 3} & \text{Mean ± SD} \\\\ \hline \text{Voxel-level (Gianchandani et al., 2024)} & 55.94 & 60.79 & 60.94 & 59.22 \pm 2.84 \\\\ \text{Global CNN} & 56.47 & 55.76 & 54.76 & 55.67 \pm 0.86 \\\\ \text{SFCN (Peng et al., 2021)} & 56.07 & 55.61 & 56.19 & 55.95 \pm 0.31 \\\\ \text{MT-ReBA with Donor} & 68.34 & \mathbf{69.97} & \mathbf{72.19} & \mathbf{70.17 \pm 1.93} \\\\ \text{MT-ReBA without Donor} & \mathbf{69.69} & 69.69 & 69.50 & 69.63 \pm 0.11 \\\\ \hline \end{array}$$ **AAL3** $$\begin{array}{lrrrr} \hline \text{Method} & \text{Seed 1} & \text{Seed 2} & \text{Seed 3} & \text{Mean ± SD} \\\\ \hline \text{Voxel-level (Gianchandani et al., 2024)} & 50.96 & 51.67 & 51.00 & 51.21 \pm 0.40 \\\\ \text{Global CNN} & 53.36 & 52.44 & 53.20 & 53.00 \pm 0.49 \\\\ \text{SFCN (Peng et al., 2021)} & 52.19 & 52.35 & 52.31 & 52.28 \pm 0.08 \\\\ \text{MT-ReBA with Donor} & 70.12 & 67.36 & 67.41 & 68.30 \pm 1.58 \\\\ \text{MT-ReBA without Donor} & \mathbf{70.74} & \mathbf{68.19} & \mathbf{68.40} & \mathbf{69.11 \pm 1.42} \\\\ \hline \end{array}$$ 2. To further rule out reliance on a single atlas, we analyzed regional age associations in 277 fully independent, held-out HCs using both HO48 and AAL3. After strict FDR correction and controlling for covariates, all 48 HO48 regions and all 166 AAL3 regions showed significant age associations. Crucially, despite their different anatomical coverages (cortical-only vs. whole-brain), both atlases revealed a highly consistent spatial pattern of age effects, with the strongest associations in limbic, insular, and frontotemporal regions. This cross-atlas consistency strongly confirms that our model captures genuine biological aging trajectories. **(p. 7, Section 4.2.3; p. 22, Appendix O)** 3. To evaluate atlas adaptability, we performed cross-atlas downstream validation in two independent cohorts (ADNI-AD and NYU-PD) using both HO48 and AAL3. The incremental benefit of regional brain-age features for AD classification was highly consistent across atlases. In ADNI, adding regional features increased AUC from 0.615 to 0.743 with HO48 ($\Delta$AUC = +0.128, 95% CI [0.065, 0.190], $p < 0.0005$), and from 0.642 to 0.762 with AAL3 ($\Delta$AUC = +0.120, 95% CI [0.061, 0.186], $p < 0.0005$). Conversely, PD classification showed disease-specific atlas sensitivity. These findings indicate that while the model captures robust aging patterns across atlases, as demonstrated in AD, the downstream discriminative utility in PD is influenced by anatomical granularity. Overall, MT-ReBA learns generalizable, whole-brain regional representations that capture anatomical heterogeneity and support the study of regional aging across different neurodegenerative diseases.**(p. 8, Section 4.2.4)**
AuthorsOpenReview ↗

Responses to Weakness 3 and Weakness 4

W3

Reviewer comment. The optimization procedure and evaluation uncertainty require further justification. The Stage 2 optimization uses separate real-image and mixed-image passes, with explicit gradient routing to prevent mixed-image losses from updating the backbone directly. While this design is understandable, the paper does not establish that separate optimization steps are necessary or superior to an implementation that combines the objectives while preserving the intended gradient restrictions. An ablation comparing these alternatives, including training stability and computational cost, would help justify the additional complexity.

Response.

The main purpose of separate optimization is to control how real and synthetic images affect gradients in different network modules, rather than to add independent training stages. Losses on real MRI update the backbone and regional branches. Losses on mixed MRI update only the regional branches, with BatchNorm statistics frozen during the mixed-image forward pass. This prevents splicing artifacts from interfering with the pretrained global age representation. We further tested the need for joint optimization through an ablation of the global and regional objectives. With only the global objective, Splice-Error was 18.95 years and SW5 was 6.0%. With only the regional objective, global MAE worsened to 24.78 years. Joint optimization achieved a global MAE of 3.10 years, a Splice-Error of 8.13 years, and an SW5 of 62.0%. This demonstrates that the two types of supervision are complementary. (p. 19, Table 9)

W4

Reviewer comment. The training dataset is modest relative to the available pretraining resources, and external generalization is not established: the pretraining dataset is rather small (<4k samples) while much larger-scale datasets are now available (i.e. OpenMind dataset with >100k samples). More importantly, the participant-level split does not establish generalization to an entirely unseen cohort. A leave-one-cohort-out evaluation or independent external validation would help determine whether the learned representations generalize across acquisition protocols and data distributions.

Response.

We added an external generalization experiment using all 561 IXI participants as a single external cohort. They were excluded from training, validation, and parameter selection and used only for the final external evaluation. This tests generalization to an unseen data source. MT-ReBA achieved an MAE of 6.681 ± 0.231 years (95% CI [6.23, 7.14]) and an RMSE of 8.666 years. These were 15.84% and 15.45% lower, respectively, than those of the next-best method, SFCN. It also achieved the highest R² (0.723) and Pearson correlation (ρ = 0.854). These results show that MT-ReBA maintains high prediction accuracy and correlation on a single external cohort, demonstrating good cross-cohort generalization. (p. 18, Appendix F)

Table 8. Generalization to a single external cohort (IXI). MAE uncertainty is based on 10,000 participant-level bootstrap resamples. MAE and RMSE are in years; RMSE, R2R^2, and ρ\rho are point estimates. Best results for each metric are shown in bold.

\hline \text{Method} & \text{MAE ± SD} & \text{95\\% CI} & \text{RMSE} & R^2 & \rho \\\\ \hline \text{y-Aware InfoNCE (Dufumier et al., 2021)} & 9.675 \pm 0.326 & [9.04, 10.32] & 12.358 & 0.438 & 0.743 \\\\ \text{SFCN (Peng et al., 2021)} & 7.938 \pm 0.272 & [7.41, 8.48] & 10.250 & 0.613 & 0.802 \\\\ \text{InfoNCE (Barbano et al., 2023)} & 10.258 \pm 0.358 & [9.57, 10.97] & 13.271 & 0.351 & 0.713 \\\\ \text{RNC (Zha et al., 2023)} & 10.390 \pm 0.341 & [9.73, 11.06] & 13.148 & 0.363 & 0.753 \\\\ \text{Triamese-ViT (Zhang et al., 2025)} & 18.667 \pm 0.572 & [17.57, 19.79] & 23.028 & -0.953 & 0.131 \\\\ \text{Anatomical} & 9.459 \pm 0.323 & [8.84, 10.10] & 12.107 & 0.460 & 0.791 \\\\ \text{MT-ReBA(Ours)} & \mathbf{6.681 \pm 0.231} & [6.23, 7.14] & \mathbf{8.666} & \mathbf{0.723} & \mathbf{0.854} \\\\ \hline \end{array}$$
Reviewer_GHy3OpenReview ↗
Rating
3(+1)
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes a method to estimated brain-age from MRI on a region-level. Usually brain age is estimated from the whole brain, which can mask different aging speeds in different parts of the brain.

Getting region-level annotations for age is not easy, and the key idea of the authors is to construct them by mixing regions from different (healthy) subjects with different ages: i.e. they swap atlas-based ROIs across subjects to obtain a new brain where some regions have a different age from the others.

Then they train a network to predict it and validate the results using neurodegenerative cohorts.

Critical strengths

The idea is sound and clever, the motivations are justified and relevant, and the results are notable. Overall this work is well positioned in the literature and I appreciate that the method itself is not overly complicated and straight to the point (I am not referring to the neural network architecture, I do not care much about that). I think this method can unlock interesting analyses.

Critical weaknesses

More important:

  • Image stitching (volume in this case) usually produces unnatural results. I see authors mention intensity matching, gaussian-feathering and soft boundaries to reduce this effect, but I do not see any results/ablations on this. Still it seems surprising that the model performs so well in downstream results given the quite simple mixing method (e.g. no adversarial reconstruction)
  • The presentation of the paper is not that easy to read. I would simplify sentence structures and make sure to introduce all concepts (e.g. donor is never defined but the abstract relies on it a lot - the typical pitfalls of LLMs prose)
  • It's not clear if the training / testing dataset mixes different sources, and whether the results are from hold-out datasets or from the same training data

Less important:

  • Reliance on a single anatomical atlas
  • The experiments on neurodegenerative cohorts should be the climax but their presentation is a bit underwhelming, e.g. there is not table, or in figure 5 there is no description of what each region is and the link with related neurodegeneration imaging literature is somehow limited. I dont see any accuracy results in AD; I would expect MT-ReBA to beat or at least be competitive with global age baseline. Figure 4 can be shortened to give more space to the results.
Questions for authors
  • I believe the loss used for prediction is L1. Why don't you start from the best method from table 1 (i.e. Barbano et al 2026) which is contrastive loss in place of L1 ?
  • Did you re-implement all baselines yourself in table 1 and table 2 ?
  • What happens if you use a different atlas ? An option, but not mandatory, could be to compare anatomical atlases vs functional atlases defined, especially for the neurodegeneration experiments.
  • Can you hold-out some datasets (e.g. ADNI) for the evaluation to assess cohort generalization ?

I am giving a conservative score because the experimental validation feels a bit preliminary, but I am willing to increase if authors provide a statisfactory response

20 comments
AuthorsOpenReview ↗

Responses to Weakness 1

Thank you for reviewing our paper and for your constructive comments. Changes are marked in the revised PDF for easy reference.

W1

Response.

In our pipeline, image stitching is used only during two-stage training as weak supervision for relative ranking. The w/o Donor setting controls whether this signal is included. We added the ablations below to test its role. All key experimental results in the paper are evaluated on unspliced, held-out test data.

  1. In Section 4.2.2, we added an ablation (w/ Donor vs. w/o Donor) to address the reviewer's concern regarding potential artificial cues from the splicing process. Our primary finding is that even after entirely removing the Donor mechanism (the simplest possible model), MT-ReBA still achieved RankAcc values of 69.63% on HO48 and 69.11% on AAL3, significantly outperforming all comparison methods. This strongly indicates that our regional brain-age prediction accuracy is driven by the core architecture and genuine structural features, rather than relying on artificial cues from image splicing. Regarding the effect of adding Donor, it acts as an auxiliary supervision signal rather than a primary performance driver. It slightly improved performance on HO48 (+0.54%) but showed a marginal decrease on AAL3 (-0.81%). This atlas-dependent fluctuation (within 1%) suggests that while the Donor mechanism serves as an effective regularization strategy for certain atlases, our model's strong performance is highly robust and does not depend entirely on this auxiliary signal.(p. 6, Section 4.2.2)

Table 2. Regional brain-age ranking accuracy (RankAcc, %). Results are mean ± SD across three training seeds.

HO48

\hline \text{Method} & \text{Seed 1} & \text{Seed 2} & \text{Seed 3} & \text{Mean ± SD} \\\\ \hline \text{Voxel-level (Gianchandani et al., 2024)} & 55.94 & 60.79 & 60.94 & 59.22 \pm 2.84 \\\\ \text{Global CNN} & 56.47 & 55.76 & 54.76 & 55.67 \pm 0.86 \\\\ \text{SFCN (Peng et al., 2021)} & 56.07 & 55.61 & 56.19 & 55.95 \pm 0.31 \\\\ \text{MT-ReBA with Donor} & 68.34 & \mathbf{69.97} & \mathbf{72.19} & \mathbf{70.17 \pm 1.93} \\\\ \text{MT-ReBA without Donor} & \mathbf{69.69} & 69.69 & 69.50 & 69.63 \pm 0.11 \\\\ \hline \end{array}$$ **AAL3** $$\begin{array}{lrrrr} \hline \text{Method} & \text{Seed 1} & \text{Seed 2} & \text{Seed 3} & \text{Mean ± SD} \\\\ \hline \text{Voxel-level (Gianchandani et al., 2024)} & 50.96 & 51.67 & 51.00 & 51.21 \pm 0.40 \\\\ \text{Global CNN} & 53.36 & 52.44 & 53.20 & 53.00 \pm 0.49 \\\\ \text{SFCN (Peng et al., 2021)} & 52.19 & 52.35 & 52.31 & 52.28 \pm 0.08 \\\\ \text{MT-ReBA with Donor} & 70.12 & 67.36 & 67.41 & 68.30 \pm 1.58 \\\\ \text{MT-ReBA without Donor} & \mathbf{70.74} & \mathbf{68.19} & \mathbf{68.40} & \mathbf{69.11 \pm 1.42} \\\\ \hline \end{array}$$ 2. In Section 4.2.6, to directly address your concern regarding whether the model merely responds to unnatural splicing artifacts, we conducted a donor-age dose–response experiment. Using the AAL3 atlas, we transplanted regions from donors of different ages into 30 healthy recipients while strictly controlling the splicing area and recipient anatomy. If the model were simply exploiting splicing artifacts, its prediction should remain constant regardless of the donor's age. However, across 74,700 pairwise comparisons (analyzed via linear mixed-effects models), MT-ReBA exhibited a robust and monotonic dose–response: a 10-year donor-age difference systematically shifted the regional predictions by 2.7–3.9 years. Crucially, the zero-dose control (donors of identical age) showed negligible prediction change (+0.05 years). These results demonstrate that while the splicing procedure is held constant, the model's predictions are driven by the biological age information of the donor tissue, rather than the splicing artifacts themselves. **(p. 9, Section 4.2.6)** 3. To directly address the possibility that the model relies on splicing artifacts, Section 4.2.3 evaluates the regional model on entirely unspliced, real-world healthy controls (HCs). If the model merely exploited artifacts introduced by the mixing method, its predictions would not generalize to unmodified data. However, in 277 held-out HCs (without splicing), the model achieved median correlation coefficients of 0.70–0.73. At the regional level, all 48 Harvard–Oxford and 166 AAL3 regions showed significant age associations (FDR $q < 0.05$). Furthermore, spatial aging patterns were consistent across both parcellations ($r = 0.589$, $p = 1 \times 10^{-5}$). These findings support that splicing serves only as auxiliary weak supervision during training, while the model learns genuine age-related regional representations.**(p. 7, Section 4.2.3; p. 22, Appendix O)**
AuthorsOpenReview ↗

Responses to Weakness 2 and Weakness 3

W2

Reviewer comment. The presentation of the paper is not that easy to read. I would simplify sentence structures and make sure to introduce all concepts (e.g. donor is never defined but the abstract relies on it a lot - the typical pitfalls of LLMs prose)

Response.

  1. We added the definition of donor to the abstract (p. 1, line 20) and Introduction (p. 2, line 61). The full definition, mathematical formulation, and implementation details are provided in Section 3.4.1 (p. 4, Section 3.4.1).

  2. A donor is a healthy participant who provides a specific atlas-defined brain parcel to a recipient during regional mixing. The donor's chronological age serves as a weak supervision signal for the regional prediction head corresponding to the replaced parcel.

W3

Reviewer comment. It's not clear if the training / testing dataset mixes different sources, and whether the results are from hold-out datasets or from the same training data

Response.

All reported results are computed on held-out data, and no subject appears in both training and test. We use three complementary protocols: 

  1. For Table 1, we constructed an internal multicenter test set of 486 HCs as follows. (p. 4, Table 1)

    • First, we grouped participants by the six data sources: GSP, ADNI, PPMI, OULU, OpenNeuro, and ICBM.

    • Within each source, we independently split participants at random into 70% training, 15% validation, and 15% test sets.

    • We then pooled the test samples from the six sources to form an independent multicenter test set of approximately 486 HCs.

  2. For Table 8, we evaluated generalization on 561 IXI participants, using IXI as an independent external cohort held out from training. (p. 18, Table 8)

  3. For Table 7, we further performed single-dataset (within-dataset) training and testing: models were trained and tested separately within ADNI and within ICBM. (p. 18, Table 7)

AuthorsOpenReview ↗

Responses to Weakness 4

W4

Reviewer comment. Reliance on a single anatomical atlas

Response.

To address your concern that MT-ReBA may depend on a single anatomical atlas, We evaluated it using two brain atlases: Harvard–Oxford (HO48) and Automated Anatomical Labeling 3 (AAL3). HO48 comprises 48 cortical regions, whereas AAL3 extends to 166 regions covering both cortical and subcortical structures, including the thalamus, substantia nigra, and locus coeruleus. This broader coverage addresses the limited subcortical coverage of HO48 and allows us to assess whether MT-ReBA captures age-related changes across the whole brain.

We added several experiments:

  1. If our model were overfitted to a specific atlas, its performance would degrade substantially when transferred. However, MT-ReBA maintained high regional age-ranking accuracy on both atlases (RankAcc: 68.30%–70.17%), consistently outperforming the best baseline methods by a large margin of 10.41–16.11 percentage points across three random seeds. Specifically, RankAcc remained robust at 69.63% on HO48 and 69.11% on AAL3.

These results strongly confirm that our model's regional ranking capability is driven by generalizable biological aging patterns, demonstrating high robustness across different anatomical resolutions and cortical/subcortical boundaries.(p. 6, Section 4.2.2)

Table 2. Regional brain-age ranking accuracy (RankAcc, %). Results are mean ± SD across three training seeds.

HO48

\hline \text{Method} & \text{Seed 1} & \text{Seed 2} & \text{Seed 3} & \text{Mean ± SD} \\\\ \hline \text{Voxel-level (Gianchandani et al., 2024)} & 55.94 & 60.79 & 60.94 & 59.22 \pm 2.84 \\\\ \text{Global CNN} & 56.47 & 55.76 & 54.76 & 55.67 \pm 0.86 \\\\ \text{SFCN (Peng et al., 2021)} & 56.07 & 55.61 & 56.19 & 55.95 \pm 0.31 \\\\ \text{MT-ReBA with Donor} & 68.34 & \mathbf{69.97} & \mathbf{72.19} & \mathbf{70.17 \pm 1.93} \\\\ \text{MT-ReBA without Donor} & \mathbf{69.69} & 69.69 & 69.50 & 69.63 \pm 0.11 \\\\ \hline \end{array}$$ **AAL3** $$\begin{array}{lrrrr} \hline \text{Method} & \text{Seed 1} & \text{Seed 2} & \text{Seed 3} & \text{Mean ± SD} \\\\ \hline \text{Voxel-level (Gianchandani et al., 2024)} & 50.96 & 51.67 & 51.00 & 51.21 \pm 0.40 \\\\ \text{Global CNN} & 53.36 & 52.44 & 53.20 & 53.00 \pm 0.49 \\\\ \text{SFCN (Peng et al., 2021)} & 52.19 & 52.35 & 52.31 & 52.28 \pm 0.08 \\\\ \text{MT-ReBA with Donor} & 70.12 & 67.36 & 67.41 & 68.30 \pm 1.58 \\\\ \text{MT-ReBA without Donor} & \mathbf{70.74} & \mathbf{68.19} & \mathbf{68.40} & \mathbf{69.11 \pm 1.42} \\\\ \hline \end{array}$$ 2. To further rule out reliance on a single atlas, we analyzed regional age associations in 277 fully independent, held-out HCs using both HO48 and AAL3. After strict FDR correction and controlling for covariates, all 48 HO48 regions and all 166 AAL3 regions showed significant age associations. Crucially, despite their different anatomical coverages (cortical-only vs. whole-brain), both atlases revealed a highly consistent spatial pattern of age effects, with the strongest associations in limbic, insular, and frontotemporal regions. This cross-atlas consistency strongly confirms that our model captures genuine biological aging trajectories. **(p. 7, Section 4.2.3; p. 22, Appendix O)** 3. To evaluate atlas adaptability, we performed cross-atlas downstream validation in two independent cohorts (ADNI-AD and NYU-PD) using both HO48 and AAL3. The incremental benefit of regional brain-age features for AD classification was highly consistent across atlases. In ADNI, adding regional features increased AUC from 0.615 to 0.743 with HO48 ($\Delta$AUC = +0.128, 95% CI [0.065, 0.190], $p < 0.0005$), and from 0.642 to 0.762 with AAL3 ($\Delta$AUC = +0.120, 95% CI [0.061, 0.186], $p < 0.0005$). Conversely, PD classification showed disease-specific atlas sensitivity. These findings indicate that while the model captures robust aging patterns across atlases, as demonstrated in AD, the downstream discriminative utility in PD is influenced by anatomical granularity. Overall, MT-ReBA learns generalizable, whole-brain regional representations that capture anatomical heterogeneity and support the study of regional aging across different neurodegenerative diseases.**(p. 8, Section 4.2.4)**
Reviewer_AnR5OpenReview ↗
Rating
4
Confidence
2
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

Given the regional aging heterogeneity in brain aging, different brain regions may follow different aging trajectories. This work, building upon previous research on regional brain aging, further focuses on the lack of ground-truth regional brain-age labels and the practical utility of regional representations. Specifically, Multi-Task Regional Brain Age Estimation (MT-ReBA) is introduced to learn region-specific brain aging representations via weak supervision and atlas-guided regional replacement, and to evaluate their utility for brain-age prediction and disease-related aging analysis. A joint global-regional framework that promotes anatomical specificity while preserving global age-prediction accuracy and a dedicated evaluation suite for regional brain aging have been carefully developed.

Critical strengths
  1. The MT-ReBA method introduced in this work is well-motivated, and the overview in Figure 1 is clear and detailed. Perhaps the typography of Figure 1 could be improved.
  2. Tables 1 and 2 both demonstrate better results compared to the previously reimplemented method. The analysis of regional anatomical specificity and the experimental details are thorough.
Critical weaknesses
  1. The atlas-guided regional replacement may rely heavily on assumptions. Perhaps donor chronological age does not necessarily reflect the true biological age of the corresponding brain region.
Questions for authors

Clarification Requests:

  1. I would like to know whether, for the medical age-estimation task, a regional age-ranking accuracy of 69.63 ± 0.11% actually has clinical value.
  2. Did the author consider that the age-related characteristics of the replaced region may not be determined solely by that region itself?
3 comments
AuthorsOpenReview ↗

Responses to Weakness 1 and Question 1

Thank you very much for taking the time to review and for your support. We try our best to address your questions as follows. Changes are marked in the revised PDF for easy reference.

W1

Reviewer comment. The atlas-guided regional replacement may rely heavily on assumptions. Perhaps donor chronological age does not necessarily reflect the true biological age of the corresponding brain region.

Response.

We use donor only as a weak supervision target matched to the source of the transplanted tissue. We demonstrate this through the following experiment.

Donor ablation. Without Donor, MT-ReBA still achieved RankAcc values of 69.63% on HO48 and 69.11% on AAL3. This shows that its regional ranking ability does not entirely depend on donor replacement during training. (p. 6, Section 4.2.2)

Table 2. Regional brain-age ranking accuracy (RankAcc, %). Results are mean ± sample SD across three training seeds.

HO48

\hline \text{Method} & \text{Seed 1} & \text{Seed 2} & \text{Seed 3} & \text{Mean ± SD} \\\\ \hline \text{Voxel-level (Gianchandani et al., 2024)} & 55.94 & 60.79 & 60.94 & 59.22 \pm 2.84 \\\\ \text{Global CNN} & 56.47 & 55.76 & 54.76 & 55.67 \pm 0.86 \\\\ \text{SFCN (Peng et al., 2021)} & 56.07 & 55.61 & 56.19 & 55.95 \pm 0.31 \\\\ \text{MT-ReBA with Donor} & 68.34 & \mathbf{69.97} & \mathbf{72.19} & \mathbf{70.17 \pm 1.93} \\\\ \text{MT-ReBA without Donor} & \mathbf{69.69} & 69.69 & 69.50 & 69.63 \pm 0.11 \\\\ \hline \end{array}$$ **AAL3** $$\begin{array}{lrrrr} \hline \text{Method} & \text{Seed 1} & \text{Seed 2} & \text{Seed 3} & \text{Mean ± SD} \\\\ \hline \text{Voxel-level (Gianchandani et al., 2024)} & 50.96 & 51.67 & 51.00 & 51.21 \pm 0.40 \\\\ \text{Global CNN} & 53.36 & 52.44 & 53.20 & 53.00 \pm 0.49 \\\\ \text{SFCN (Peng et al., 2021)} & 52.19 & 52.35 & 52.31 & 52.28 \pm 0.08 \\\\ \text{MT-ReBA with Donor} & 70.12 & 67.36 & 67.41 & 68.30 \pm 1.58 \\\\ \text{MT-ReBA without Donor} & \mathbf{70.74} & \mathbf{68.19} & \mathbf{68.40} & \mathbf{69.11 \pm 1.42} \\\\ \hline \end{array}$$ **Donor-age dose–response.** We held the anatomy of 30 recipients fixed and analyzed 74,700 pairwise donor comparisons. Regional predictions changed monotonically with donor-age differences, with test-set slopes of 0.267–0.334. The zero-dose control showed no significant shift. This supports an age-related response to the donor tissue, beyond a generic response to tissue replacement. **(p. 9, Section 4.2.6)** **Independent validation on real MRI.** In 277 independent HCs with unspliced scans, all 48 HO48 regions and all 166 AAL3 regions showed significant associations between predicted regional age and chronological age after FDR correction. This shows that the model captures regional information related to natural aging on real MRI. **(p. 7, Section 4.2.3; p. 22, Appendix O)** Together, these findings support donor age as an informative weak supervision signal for learning anatomically localized, age-related representations. ### Q1 **Reviewer comment.** I would like to know whether, for the medical age-estimation task, a regional age-ranking accuracy of 69.63 ± 0.11% actually has clinical value. **Response.** We added neurodegenerative disease experiments to assess the practical value of MT-ReBA through disease classification and disease-related regional differences. Adding regional BAGs improved AD classification. Under HO48, AUC increased from 0.615 to 0.743 (reported $\Delta\mathrm{AUC}=+0.128$, 95% CI [0.065, 0.190]). Under AAL3, it increased from 0.642 to 0.762 (reported $\Delta\mathrm{AUC}=+0.120$, 95% CI [0.061, 0.186]). Both gains were significant ($p < 0.0005$). The regional analysis also identified disease-related anatomical patterns. AD effects were strongest in the medial temporal lobe, including the posterior parahippocampal cortex (+10.93 years). PD effects were mainly found in motor-related regions, including the supplementary motor area (+6.48 years) and precentral gyrus (+6.04 years). These effects passed FDR correction. These findings show that regional brain age adds information beyond global brain age for AD classification. It also reveals regional changes with disease-related anatomical distributions. **(pp. 8–9, Sections 4.2.4–4.2.5)**
AuthorsOpenReview ↗

Responses to Question 2

Q2

Reviewer comment. Did the author consider that the age-related characteristics of the replaced region may not be determined solely by that region itself?

Response.

We thank the reviewer for this insightful question. We fully acknowledge that in a biological brain, a region's characteristics are influenced by its network context. To directly test whether our model can isolate local age information from global contextual interference, we designed two stringent controls:

1. Isolating regional age while strictly maintaining global context:

In the donor-age dose-response experiment (Section 4.2.6), we held the recipient's entire brain anatomy fixed. This means the global network context and all other brain regions were completely unchanged. The only variable was the age of the transplanted donor tissue. If the model's prediction shifts were merely an artifact of disrupted global context, the shift would be constant regardless of donor age. However, MT-ReBA showed a clear dose-response (slopes 0.267–0.334) to the donor's age, while the zero-dose control (identical age, no contextual mismatch) showed near-zero shift. This proves the model can extract age information strictly from the local region itself, independent of the global context.

2. Proving anatomical specificity via occlusion:

To further rule out the model relying on a global "mismatch" signal, we conducted ROI occlusion on unspliced, original MRI (Appendix N). If the model were just responding to holistic context, occluding any region would cause broad, unspecific prediction changes. Instead, we found that occluding a specific region predominantly altered its corresponding regional prediction (response ratio: 14.45 times, localization accuracy: 75.0%, chance level 2.1%, p=0.0005p = 0.0005). This confirms that the regional heads are highly localized and do not drift with global anatomical context.

Summary: While the reviewer rightly points out the interconnected nature of the brain, our strict control-variable experiments demonstrate that MT-ReBA successfully disentangles and relies on identifiable local age-related features, rather than merely reacting to disrupted global context. (pp. 9–10, Section 4.2.6; pp. 21–22, Appendix N)

Reviewer_AnR5OpenReview ↗

Official Comment by Reviewer AnR5

Dear Authors,

Thank you for your further responses; they really help me and readers to further understand your work.

Best,
Reviewer AnR5

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.