Abstract

Brain age gap (BAG) derived from structural MRI is a promising surrogate biomarker for quantifying neurodegenerative burden across the Alzheimer’s disease (AD) continuum; however, most existing methods emphasize global accuracy while offering limited support for region-level interpretation. We propose VCDA-Net, a multi-scale brain age estimation framework that couples global morphological context with local geometric cues derived from voxel-cluster density. VCDA-Net introduces Voxel-Cluster Density Attention (VCDA) to encode stable regional geometry from high-saliency features and models inter-regional relations via a multi-attributed brain graph with Multi-Edge Attention. A Transformer aggregation module captures long-range dependencies across 32 anatomical structures, while an adaptive gated fusion integrates graph-derived signatures with U-Net bottleneck priors for final age regression and structure-wise BAG profiling. Evaluated on multi-cohort normative modeling and the AD-spectrum, VCDA-Net achieves an MAE of 3.71 ± 0.31 years (r = 0.97, R^2 = 0.94). Our results reveal progressively increased BAG from MCI to AD, with regional BAG highlighting AD-vulnerable structures such as the bilateral hippocampus and the ventricular system.

Links to Paper and Supplementary Materials

Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/3526_paper.pdf

SharedIt Link: Not yet available

SpringerLink (DOI): Not yet available

Supplementary Material: Not Submitted

Link to the Code Repository

https://github.com/quoctienchonnam25-collab/Voxel-Cluster_Density_Attention_Network

Link to the Dataset(s)

ABIDE dataset: https://ida.loni.usc.edu/login.jsp ADNI dataset: https://ida.loni.usc.edu/login.jsp PPMI dataset: https://ida.loni.usc.edu/login.jsp IXI dataset: https://brain-development.org/ixi-dataset/ OASIS-1 dataset: https://sites.wustl.edu/oasisbrains/home/oasis-1/ OASIS-2 dataset: https://sites.wustl.edu/oasisbrains/home/oasis-2/

BibTex

@InProceedings{NguQuo_VCDANet_MICCAI2026,
        author = { Nguyen, Quoc-Tien AND Lee, Myungeun AND Kim, Soo-Hyung AND Yang, Hyung-Jeong},
        title = { { VCDA-Net: Graph-Guided Voxel-Cluster Density Attention for Global-Regional Brain Age Gap Estimation } },
        booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
        year = {2026},
        publisher = {Springer Nature Switzerland},
        volume = {LNCS 16894},
        month = {September},
        page = {pending}
}


Reviews

Review #1

  • Please describe the contribution of the paper

    This paper proposes VCDA-Net, a multi-scale brain age estimation framework that balances global predictive accuracy with region-level interpretability. By coupling global morphological context with local geometric cues via a novel Voxel-Cluster Density Attention (VCDA) mechanism , the model captures both holistic and fine-grained neurodegeneration. Integrated with a multi-attributed brain graph and Transformer aggregation , it achieves high predictive fidelity and enables structure-wise Brain Age Gap (BAG) profiling to quantify neurodegenerative burden across the Alzheimer’s disease continuum.

  • Please list the major strengths of the paper: you should highlight a novel formulation, an original way to use data, demonstration of clinical feasibility, a novel application, a particularly strong evaluation, or anything else that is a strong aspect of this work. Please provide details, for instance, if a method is novel, explain what aspect is novel and why this is interesting.

    1.The core idea of coupling global morphological context with local geometric cues via the novel Voxel-Cluster Density Attention (VCDA) mechanism is highly interesting. This approach effectively addresses the limitations of traditional regional volume metrics by mitigating inherent segmentation noise and capturing stable, fine-grained structural variations. 2.The quantitative evaluation is solid and persuasive. The proposed VCDA-Net demonstrates competitive performance by outperforming several established baselines (including SFCN, ResNet50, and DenseNet121), achieving a mean absolute error (MAE) of 3.71years on a multi-cohort normative dataset. 3.The methodology thoughtfully incorporates a linear bias correction step to mitigate the well-known regression-to-the-mean effect in age estimation. This crucial adjustment not only improves the overall accuracy of the brain age predictions but also ensures that the resulting Brain Age Gap (BAG) reliably isolates true neurodegenerative changes from systematic prediction errors.

  • Please list the major weaknesses of the paper. Please provide details: for instance, if you state that a formulation, way of using data, demonstration of clinical feasibility, or application is not novel, then you must provide specific references to prior work.

    1.The review of related work in the Introduction could be better structured. The current narrative feels slightly disjointed. It would be highly beneficial to reorganize this section to explicitly tighten the connection between existing multimodal learning strategies, global-local brain age prediction paradigms, and the specific methodological gap that VCDA-Net aims to fill. 2.The multi-attributed graph utilizes three specific metrics for edge construction (dot product, Euclidean distance, and cosine similarity). The manuscript lacks a clear justification for selecting these exact three. Given that they measure overlapping concepts of geometric similarity, there is a high risk of information redundancy. Furthermore, the ablation study (Table 2) only evaluates individual metrics against the full combination; it lacks experiments evaluating pairwise combinations (e. g. , using only two metrics) to definitively prove that all three are strictly necessary and non-redundant. 3.While the evaluation is extensive, several of the primary comparison baselines (e. g. , ResNet50, DenseNet121, and SFCN) are relatively dated. To robustly validate the framework as a state-of-the-art approach, the authors should include comparisons with more recent and highly representative architectures in this domain, such as Global-Local Transformer-based brain age models. 4.The authors are strongly advised to double-check the statistical values reported in Table 3.Some of the mean and standard deviation values for the clinical groups (e. g. , specifically the magnitude of standard deviations relative to the means) warrant a careful verification to ensure there are no typographical errors in the calculations or reporting. 5.A fundamental aspect of the methodology requires further clarification: how is the “ground truth” age for individual anatomical regions determined? While chronological age serves as a global label, the manuscript does not explicitly detail the biological or mathematical assumptions used to assign or approximate target ages for specific local brain structures during training, especially concerning the physiological differences between healthy controls and pathological subjects.

  • Please rate the clarity and organization of this paper

    Satisfactory

  • Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.

    The submission has provided an anonymized link to the source code, dataset, or any other dependencies.

  • Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?

    N/A

  • Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2026/en/REVIEWER-GUIDELINES.html

    1.While the authors emphasize the necessity of the linear bias correction to mitigate the regression-to-the-mean effect (Section 2.4) and demonstrate its effect on aligning the healthy control mean BAG to zero (Section 3.3), there is no explicit ablation study evaluating its quantitative impact. It would strengthen the paper to report how the bias correction step quantitatively affects the main predictive metrics (e. g. , overall MAE/RMSE) and, more importantly, whether the statistical significance of differentiating AD/MCI from healthy controls changes strictly as a result of applying this correction. 2.While the model achieves strong performance on cross-sectional baseline scans, the true clinical utility of the Brain Age Gap (BAG) as a progressive biomarker relies heavily on its longitudinal stability and sensitivity to disease progression. Given that cohorts like ADNI and OASIS contain rich longitudinal data (which the authors explicitly restricted to baseline scans), it is highly recommended to evaluate the framework on longitudinal trajectories. Demonstrating that the subject-specific regional BAG predictably worsens as a patient transitions from MCI to AD over several years would substantially elevate the clinical impact and validity of this work.

  • Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.

    (3) Weak Reject — marginally below the acceptance threshold, but would not mind if accepted, dependent on rebuttal

  • Please justify your recommendation. What were the major factors that led you to your overall score for this paper?

    While the coupling of global morphological context with local voxel-cluster density is an interesting concept, the manuscript lacks critical justifications. Specifically, the selection of the three edge attribute components (dot product, Euclidean distance, and cosine similarity) is not explained, raising concerns regarding feature redundancy. Furthermore, the lack of an explicit ablation study for the age bias correction step makes it difficult to assess the true predictive gain of the architecture versus statistical post-processing.Additionally, the comparison baselines are somewhat dated (e.g., ResNet50, SFCN). Finally, the lack of clarity regarding how regional “ground truth” labels are determined.

  • Reviewer confidence

    Confident but not absolutely certain (3)

  • [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.

    Accept

  • [Post rebuttal] Please justify your final decision from above.

    N/A



Review #2

  • Please describe the contribution of the paper

    The authors propose VCDA-Net, which is a brain age gap estimation network that provides both global and regional brain age predictions. The proposed approach uses a graph-guided voxel-cluster attention mechanism to improve its results compared to the literature. The proposed mechanism seems novel compared to existing works in the literature (SFCN, UNet, etc.)

  • Please list the major strengths of the paper: you should highlight a novel formulation, an original way to use data, demonstration of clinical feasibility, a novel application, a particularly strong evaluation, or anything else that is a strong aspect of this work. Please provide details, for instance, if a method is novel, explain what aspect is novel and why this is interesting.

    The main strength of the paper is its novel methodology, which uses a graph-guided voxel-cluster density attention mechanism to improve brain age gap estimates. The guidance seems to improve age gam predictions, and the fact that it relies on voxel-clusters instead of segmentations masks potentially improves the quality of the regional brain age predictions.

  • Please list the major weaknesses of the paper. Please provide details: for instance, if you state that a formulation, way of using data, demonstration of clinical feasibility, or application is not novel, then you must provide specific references to prior work.

    The main weakness of the paper is that its methodology is hard to follow based solely on the manuscript. Fortunately, the authors provided the anonymized code. For example, in Figure 1 (a), where the regional brain age is coming from? It appears in Figure 1(b) (“regional brain age estimator”), but it is unclear where it fits within VCDA-Net workflow.

    • I missed a figure showing the regional age predictions for the MCI and ADS groups.

    • Were the models that VDCA-Net compared against retrained on the same data?

    • I believe the data has sex and not gender information.

  • Please rate the clarity and organization of this paper

    Good

  • Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.

    The submission has provided an anonymized link to the source code, dataset, or any other dependencies.

  • Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?

    N/A

  • Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2026/en/REVIEWER-GUIDELINES.html

    N/A

  • Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.

    (3) Weak Reject — marginally below the acceptance threshold, but would not mind if accepted, dependent on rebuttal

  • Please justify your recommendation. What were the major factors that led you to your overall score for this paper?

    The paper addresses the important problem of brain-age gap prediction at both global and regional levels. The proposed model appears to improve results relative to the literature. However, the explanation of the proposed methodology is hard to follow and needs improvement. The regional age results are inconsistent. Why does the brain age gap appear negative for many brain regions (Figure 3 - top)?

  • Reviewer confidence

    Confident but not absolutely certain (3)

  • [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.

    Accept

  • [Post rebuttal] Please justify your final decision from above.

    The authors addressed most of my concerns and clarified the methodology, which was one of my main concerns.



Review #3

  • Please describe the contribution of the paper

    1.The paper introduces the VCDA mechanism, which encodes stable regional geometry by identifying high-saliency features. 2.The proposed method constructs a multi-attributed brain graph comprising 32 anatomical structure nodes, utilizing a Multi-Edge Attention mechanism and a Transformer aggregation module to capture long-range regional dependencies. 3.The proposed method implements structure-wise BAG profiling, which not only predicts global brain age but also localizes specific neuroanatomical structures vulnerable to AD.

  • Please list the major strengths of the paper: you should highlight a novel formulation, an original way to use data, demonstration of clinical feasibility, a novel application, a particularly strong evaluation, or anything else that is a strong aspect of this work. Please provide details, for instance, if a method is novel, explain what aspect is novel and why this is interesting.

    1.The VCDA mechanism captures local morphological details, whereas existing DL models often ignore heterogeneous information within specific structures. 2.The proposed method is able to localize neurodegenerative changes to specific anatomical structures (e.g., hippocampus and ventricles in AD), greatly enhancing method interpretability.

  • Please list the major weaknesses of the paper. Please provide details: for instance, if you state that a formulation, way of using data, demonstration of clinical feasibility, or application is not novel, then you must provide specific references to prior work.

    1.While the paper utilizes a large-scale dataset containing 1,710 MRI scans, there is a lack of an unseen dataset for validation, and the generalization of the model needs to be verified in future studies. 2.More decimal places are needed in Table 1 to differentiate the R2 and r values among methods.

  • Please rate the clarity and organization of this paper

    Good

  • Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.

    The submission has provided an anonymized link to the source code, dataset, or any other dependencies.

  • Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?

    N/A

  • Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2026/en/REVIEWER-GUIDELINES.html

    N/A

  • Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.

    (5) Accept — should be accepted, independent of rebuttal

  • Please justify your recommendation. What were the major factors that led you to your overall score for this paper?

    The proposed method captures local morphological details and is able to localize neurodegenerative changes to specific anatomical structures, which greatly enhances model interpretability.

  • Reviewer confidence

    Very confident (4)

  • [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.

    Accept

  • [Post rebuttal] Please justify your final decision from above.

    The VCDA mechanism proposed in this paper is novel and the evaluation on 1,710 MRI scans is solid.



Author Feedback

We thank all reviewers for their constructive feedback and for recognizing the strengths of our work, including the “novel methodology” with a “highly interesting mechanism” (@R1, R2), the effective “voxel-cluster density attention” for “capturing local morphological details” and “greatly enhancing interpretability” (@R2,R3), and the “crucial adjustment” of bias correction to “reliably isolate true neurodegenerative changes”(@R1). Q1.Non-redundancy(@R1): Cosine Similarity, Euclidean Distance, and Dot Product capture complementary geometric perspectives: orientation, magnitude, and their composite relationship. Orientation identifies shape patterns, while magnitude captures scale-specific morphological changes. To prevent redundancy, Gated Fusion dynamically filters these inputs, ensuring only non-overlapping features are integrated. Table 2 proves this synergy: single-metric models using Euclidean (MAE:4.42), Dot product (MAE:4.40), or Cosine (MAE:5.82) underperform VCDA-Net (MAE:4.22). This MAE reduction confirms that multi-edge attributes provide unique information that Gated Fusion prioritizes, validating the framework’s necessity for capturing complex structural variations. We agree that pairwise ablation studies would further highlight the specific contribution of each metric pair. Q2.Statistical Verification(@R1,R2): We confirmed the accuracy of the reported Mean+/- SD values. The high variance reflects inherent biological heterogeneity typical in multi-site neuroimaging. Crucially, in Table 3, the 95%CIs for clinical groups (e.g., ADNI AD [3.06, 5.11]) remain clearly distinct from zero and the HC group [-0.74,0.74], confirming that observed BAG increases are statistically significant indicators of pathology. Furthermore, negative BAG values in Fig. 3 represent normal biological diversity; by centering the HC mean at zero, these values identify brain regions that appear younger than their chronological age. Q3.Assumptions for Regional Ground Truth(@R1,R2): We clarify that chronological age serves as the indirect ground truth for all 32 regional decoders (Eq.3,4), based on the established assumption that healthy regional development aligns with chronological time. By fitting our accelerated aging analysis framework exclusively on Healthy Controls, we establish a normative reference that effectively decouples systematic prediction bias from true pathological signals. As detailed in Section 3.3 (p.8), our results confirm that this strategy successfully isolates neurodegeneration: despite training on global labels, VCDA-Net identifies significant localized deviations in AD groups, particularly in the Right Hippocampus (BAG=2.11±8.08,p<.005) and Right Lateral Ventricle (BAG=3.94±7.39, p<.001). These gaps quantify region-specific “accelerated aging” as deviations from the normative trajectory, providing a robust proxy in the absence of regional biological labels. Q4.SOTA Comparison & Impact(@R1): VCDA-Net was benchmarked against established 3D CNNs (SFCN, ResNet) and modern Transformer-based SOTA (Voxel-level BA using Swin UNETR) alongside specialized frameworks (SynthBA, Ordinal-DR), covering both gold standards and the global-local paradigms requested. The bias correction is integral; its efficacy is evidenced by significant AD findings (Sec 3.3). We value the suggestion on longitudinal trajectories; while this study builds a robust cross-sectional foundation, extending the framework to track disease progression is a key priority for future development. Q5.Refinement & Generalization(@R1,R2,R3): We will slightly refine the Introduction and update technical terms (e.g., changing “Gender” to “Sex”) and Table 1 precision as suggested. Regarding generalization, our multi-site dataset (1,710 scans) provides robust evidence of model stability. While all benchmarks were evaluated under identical conditions for fair comparison, we agree that validation on an unseen dataset is a valuable direction for future expansion.




Meta-Review

Meta-review #1

  • Your recommendation

    Invite for Rebuttal

  • Please justify your decision. In case you deviate from the reviewers’ recommendations, explain in detail the reasons why. In case of an invitation for rebuttal, clarify which points are important to address in the rebuttal.

    The reviews stated different recommendations, with R1 and R2 recommending weak reject and R3 recommending accept. All reviewers raised concerns about the paper that should be addressed in the rebuttal. The major points addressed by the reviewers include clarifying the regional brain age pipeline and its supervision signal, addressing comparison fairness, justifying the choice of metric, and addressing and stating any other missing information requested by the reviewers.

  • After you have reviewed the rebuttal and updated reviews, please provide your recommendation based on all reviews and the authors’ rebuttal.

    Accept

  • Please justify your recommendation.

    Following the rebuttal, all three reviewers agreed to accept the paper, after the authors clarified the regional brain age pipeline, the supervision signa and the methodology, which R1 and R2 had initially found insufficiently explained.



Meta-review #2

  • After you have reviewed the rebuttal and updated reviews, please provide your recommendation based on all reviews and the authors’ rebuttal.

    Accept

  • Please justify your recommendation.

    The rebuttal successfully addressed most methodological concerns, resulting in post-rebuttal support for acceptance by all reviewers and leading to my recommendation of acceptance.



Meta-review #3

  • After you have reviewed the rebuttal and updated reviews, please provide your recommendation based on all reviews and the authors’ rebuttal.

    Accept

  • Please justify your recommendation.

    This paper introduces a multi-scale brain age estimation framework that utilizes a graph-guided voxel-cluster density attention mechanism for both global and regional predictions. In the rebuttal, the authors clarified that chronological age serves as an indirect ground truth for regional decoders to establish a normative reference. They also justified the complementary nature of the chosen geometric metrics and the statistical validity of their results. Convinced by these clarifications, both Reviewers 1 and 2 upgraded their scores, leading to a unanimous agreement that the paper is a strong methodological contribution suitable for MICCAI.



back to top