List of Papers Browse by Subject Areas Author List
Abstract
Accurate characterization of brain aging is essential for understanding cognitive decline and assessing the risk of neurodegenerative disease. Brain age estimated from cross-sectional magnetic resonance imaging (MRI) data provides a snapshot of brain health relative to chronological age, and the derived brain age delta has emerged as a promising biomarker. However, brain age delta reflects cumulative effects at a single time point and fails to capture ongoing aging dynamics. Longitudinal approaches address this limitation by estimating age differences between scan pairs to derive aging speed. Nevertheless, existing methods primarily rely on intensity or texture differences between images, overlook the spatial heterogeneity of aging processes, and provide limited interpretability. To overcome these limitations, we propose a novel framework that estimates brain aging speed from longitudinal deformation fields obtained via diffeomorphic image registration. Instead of solely generating a single global estimate, our method produces patch-wise local aging predictions and adaptively integrates them into a unified global prediction, improving both predictive performance and interpretability. Evaluated on large-scale datasets, our approach achieves superior accuracy compared with existing methods and enhances the identification of abnormal aging patterns in diseased populations. Code is available at https://github.com/Kateridge/MorphologyAgingSpeed.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/3925_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to the Code Repository
https://github.com/Kateridge/MorphologyAgingSpeed
Link to the Dataset(s)
ADNI dataset: https://adni.loni.usc.edu/data-samples/adni-data/
OASIS-3 dataset: https://sites.wustl.edu/oasisbrains/home/oasis-3/
AIBL: https://aibl.org.au/
BibTex
@InProceedings{ZhaYua_Interpretable_MICCAI2026,
author = { Zhang, Yuanwang AND Li, Hongming AND Fan, Yong},
title = { { Interpretable Local-to-Global Estimation of Brain Aging Speed From Morphological Changes Using Longitudinal Structural MRI Data } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 16887},
month = {September},
page = {pending}
}
Reviews
Review #1
- Please describe the contribution of the paper
The paper proposes a longitudinal framework to estimate brain aging speed using deformation fields derived from diffeomorphic registration between MRI scan pairs. Instead of relying on intensity differences, it models spatially heterogeneous aging by producing patch-wise local aging estimates across the brain. These local predictions are then adaptively aggregated into a global aging speed, improving interpretability and capturing regional dynamics. Experiments on ADNI, OASIS3 and AIBL datasets show improved accuracy and better detection of abnormal aging patterns compared to prior methods.
- 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 approach is novel and interesting to model longitudinal morphological trajectories based on a diffeomorphic registration framework.
- The results are strong, especially for cross-site generalization compared to previous state-of-the-art.
- 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 first part of your model (deep learning-based diffeomorphic image registration) follows exactly the VoxelMorph pipeline. Is it trained along with the actual regressor of aging speed based on the Jacobian determinant map? If yes, do you observe a correlation between the registration accuracy and the speed aging prediction accuracy?
- (Suggestion) In your model, you only consider time difference (t2-t1) between two scans, implicitly assuming the morphological changes do not depend on the initial timepoint t1 (i. e. original age). According to current brain charts [Bethlehem, Nature 2022], we know this is not true because brain morphology typically evolves non-linearly with time (and is even different across sex). I wonder whether including the original timepoint t1 (and maybe sex) as input to the model (in addition to the two scans) would refine the quality of the brain aging speed estimation.
- 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 authors claimed to release the source code and/or dataset upon acceptance of the submission.
- 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?
This paper proposes an elegant framework to model regional brain aging speed based on diffeomorphic fields. The model is simple and the results are strong compared to prior works. In practice, it allows to accurately distinguish MCI -> AD converters from non-converters using MRI at the group level.
- 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 have successfully answered my main concern about age- and sex-dependent modeling of brain atrophy. I believe there is still room for improvement but the results given here are already promising and deserve to be published at MICCAI.
Review #2
- Please describe the contribution of the paper
The authors propose a method to predict age differences between two longitudinal brain MRI scans by leveraging morphological information derived from diffeomorphic registration, combined with patch-based localized estimation. The experiments suggest that using morphology-based features improves performance compared to relying solely on intensity information, and that localized (patch-level) modeling outperforms global representations.
- 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.
- Well-motivated and clearly presented idea: The paper is built on a clear and intuitive premise that longitudinal data can better capture aging dynamics compared to cross-sectional approaches. In addition, the use of localized (patch-level) information is well justified, as it enhances both interpretability and predictive performance.
- Meaningful use of morphological information: Leveraging morphology derived from diffeomorphic registration is a strong design choice, as it focuses on biologically relevant changes rather than relying solely on intensity-based features.
- Comprehensive experimental validation: The experiments are generally sufficient to support the main claims, including comparisons between morphology vs. intensity features and localized vs. global modeling.
- Evaluation on downstream clinical tasks: The inclusion of experiments on diseased populations strengthens the practical relevance of the work and demonstrates potential clinical utility.
- 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.
- Limited methodological novelty: While the combination of morphology-based features and patch-level modeling is well motivated, each component has been explored in prior work (e.g., registration-based morphometry and patch-based learning). The paper would benefit from a clearer articulation of what is fundamentally new beyond this integration, and how it advances the state of the art.
- Missing details on hyperparameter weighting: The paper does not specify the hyperparameters that control the relative strength of each loss term.
- Unexpected generalization behavior: The model achieves lower error (MAE and RMSE) on the external test set than the internal set, which is counterintuitive.
- 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 authors claimed to release the source code and/or dataset upon acceptance of the submission.
- 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.
(4) Weak Accept — marginally above the acceptance threshold, but would not mind if rejected, dependent on rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
The paper presents a well-motivated approach to longitudinal brain age modeling, and the experimental results generally support the main claims. The use of morphology and localized estimation is meaningful and relevant. However, limitations such as modest novelty, missing methodological details (e.g., hyperparameters) weaken the overall contribution. Overall, I find the work promising but with some shortcomings, leading to a weak accept recommendation.
- Reviewer confidence
Confident but not absolutely certain (3)
- [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.
N/A
- [Post rebuttal] Please justify your final decision from above.
N/A
Review #3
- Please describe the contribution of the paper
The paper proposes a framework for estimating brain aging speed from longitudinal structural MRI by leveraging deformation fields obtained via diffeomorphic registration.
Instead of producing a single global prediction, the method: -computes Jacobian determinant maps to quantify voxel-wise morphological changes, -partitions these maps into patches, -predicts local aging speed per patch, -aggregates local predictions into a global estimate using learned relevance weights.
The framework is trained on cognitively unimpaired subjects and evaluated on longitudinal datasets, with the goal of improving both predictive performance and interpretability compared to existing longitudinal and cross-sectional approaches.
- 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 paper addresses an important limitation of brain age delta by explicitly modeling aging dynamics through longitudinal data, which is a meaningful and relevant direction.
2.The use of deformation fields and Jacobian determinants as representations of structural change is technically sound and biologically meaningful compared to intensity-based differences.
3.The patch-wise regression framework is an interesting design choice that enables spatial interpretability while maintaining a global prediction.
4.The paper includes: -comparisons with cross-sectional and longitudinal baselines, -ablation studies (without morphology / without local prediction), -evaluation on internal and external datasets.
- 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.Figure 3 presents trends of aging speed across different diagnostic progression groups (e.g., CU-to-CU, MCI-to-AD), computed as averages within age bins. However, the manuscript does not report: -the number of samples per group -the number of samples per age bin -any measure of variability (e.g., standard deviation, confidence intervals, or error bars) This is a critical limitation, as the observed trends may be strongly influenced by differences in sample size across groups or bins. In particular, progression groups are likely smaller and unevenly distributed across age ranges, which can lead to unstable estimates and misleading trends. As a result, it is difficult to determine whether the reported differences reflect genuine biological effects or are artifacts of sampling variability.
2.The paper concludes that progression groups exhibit higher aging speed based on the trends in Figure 3.However, no statistical test is reported to compare these groups, and no uncertainty estimates are provided. Given the lack of quantitative support, these conclusions appear overstated.
3.The proposed aging speed metric is central to the contribution, yet its validity is not rigorously demonstrated. In particular, the paper does not show: -correlation with clinical outcomes or cognitive measures, -comparison with established biomarkers (e.g., brain age delta) beyond qualitative trends, -robustness or stability of the metric across subjects.
As a result, it remains unclear whether the proposed measure captures meaningful biological aging dynamics or primarily reflects model-specific behavior.
4.The method relies entirely on deformation fields obtained via diffeomorphic registration. However, no analysis is provided regarding: -sensitivity to registration errors, -robustness across different registration approaches, -potential artifacts introduced by misalignment.
Given that the Jacobian determinant maps are the core input to the model, this is an important missing validation.
- 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 authors claimed to release the source code and/or dataset upon acceptance of the submission.
- 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 proposes a well-designed framework for estimating brain aging speed from longitudinal MRI data and demonstrates improved predictive performance over existing methods.
However, the evaluation does not provide sufficient statistical support for key claims, particularly regarding differences between diagnostic groups. The absence of information on sample sizes per group and per age bin, as well as the lack of uncertainty estimates, makes it difficult to assess the reliability of the reported trends.
In addition, the proposed aging speed metric is not thoroughly validated, and the method’s dependence on image registration is not analyzed in terms of robustness.
Overall, while the method is promising, the current evidence is not sufficient to fully support the conclusions, placing the paper slightly below the acceptance threshold.
- 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.
Thank you for the detailed and well-structured rebuttal. I appreciate the authors’ efforts in directly addressing the main concerns raised in the review.
In particular, the additional clarification regarding Figure 3, including the reporting of sample sizes per group and age bin, as well as the inclusion of statistical testing, significantly strengthens the validity of the group-level analysis. This resolves my primary concern regarding the interpretability of the reported trends.
Furthermore, the additional validation of the proposed aging speed metric through statistically significant correlations with cognitive measures (e.g., MMSE, CDR, ADAS13) provides important evidence supporting its clinical relevance, addressing another key limitation identified in the original review.
While the discussion on registration robustness remains somewhat qualitative, the authors provide reasonable justification given the within-subject setting and the scope of the rebuttal.
Overall, I believe the rebuttal has sufficiently addressed the major concerns raised in my review. I therefore update my recommendation to acceptance.
Author Feedback
We sincerely thank all reviewers for their constructive feedback. Following meta-review, we first address the methodological and statistical concerns, then respond to additional points.
Q1: Statistical rigor of group comparisons (Fig. 3) [R4-C1,C2; Meta-Review] We appreciate this important concern. We have amended Fig. 3 to include 1) sample sizes per group and per age bin, and 2) Wilcoxon rank-sum test results between converter and non-converter groups, overall and within each age bin. Brain age delta showed significant differences (p < 0.01) between two non-converter groups of different disease severity (CU->CU vs. MCI->MCI) across all age bins and overall. For aging speed, MCI->AD differed significantly from both non-converter groups across all age bins (p < 0.01), while CN->MCI differed significantly in the 60-64 age bin. Crucially, the non-converter groups (CU->CU vs. MCI->MCI) showed no significant aging speed differences across most age bins, confirming that aging speed captures disease progression rather than severity, unlike brain age delta.
Q2: Age and sex as confounders [R1-C2; Meta-Review] We agree brain morphology changes nonlinearly with age and differs between sex. In our framework, the Jacobian determinant maps may already implicitly encode age-dependent atrophy patterns. Incorporating t_1 as an input feature at the output layer of regressors yielded comparable performance on CU subjects (MAE :1.15, R2: 0.67 internal test; MAE: 0.72, R2: 0.58 external test), suggesting Jacobian maps already capture substantial age-related information. However, more advanced conditioning methods (e.g., FT-Transformers or FiLM) may further improve predictions, which is a promising future work, and we will discuss this limitation in the revised manuscript.
Q3: Validation of the aging speed metric [R4-C3] Spearman correlations between the estimated aging speed and longitudinal cognitive change (on ADNI CU test + all ADNI MCI/AD subjects) were statistically significant: MMSE (-0.23), CDR (0.28), FAQ (0.24), ADAS13 (0.23), all p < 0.001.These further confirm aging speed captures clinically meaningful cognitive decline, not merely model-specific behavior.
Q4: Registration robustness [R1-C1, R4-C4] Our framework is end-to-end and jointly trained; registration and regression errors decrease simultaneously, with registration converging first. The registration part is a within-subject registration task, which is inherently easier than cross-subject registration, reducing susceptibility to large misalignment. Also, Table 2 (rows 5–6 vs. 2–3) consistently shows improvements over methods without morphological features, demonstrating the utility and robustness of deformation-based representations. Nevertheless, we agree that a systematic comparison across registration methods is a valuable future work.
Q5: Novelty [R3-C1] Our contribution addresses two limitations of prior works: 1) intensity-based methods are sensitive to imaging noise and variability, whereas within-subject registration-derived Jacobian determinants provide morphology-driven, noise-robust features reflecting anatomical aging; 2) existing methods predict a single global aging speed, whereas our patch-wise prediction with learned aggregation explicitly models spatial heterogeneity, improving both prediction accuracy (Table 2) and interpretability (Fig. 4). Their integration within a jointly optimized framework is, to our knowledge, novel.
Q6: Hyperparameter details [R3-C2] Default hyperparameters were used for registration, similar to prior studies [5]. The full loss is: 1.0L_reg + 0.01L_smo + 1.0L_local + 1.0L_global. These details will be added to the manuscript.
Q7: Lower external test error [R3-C3] This reflects dataset characteristics: AIBL has a narrower age range and smaller longitudinal intervals (\Delta t) than ADNI/OASIS (Table 1). All comparison methods achieved lower external errors, confirming a dataset-level rather than model-specific effect.
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 paper got mostly positive feedback. Still, there are important methodological points raised by Rev. 4 that need to be addressed in a rebuttal. An additional point to discuss, raised by Rev. 1, is that the speed is assumed in the paper to be independent of age and sex, which is not true. Thus, the method is expected to fail when applied to younger populations.
- 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 authors addressed the reviewers’ main concerns. Brain aging speed is a useful complement to brain aging estimation, providing additional features to characterize diseased populations.
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.
I think that the idea is interesting. Although some weaknesses still remain, I thought that authors have provided a good discussion on this early-stage but interesting research, and have provided some convincing responses to specific critiques. I strongly encourage authors to include better clarifications and revisions as suggested by reviewers to strengthen the final version. I would argue for accept.
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.
Although pairwise training has been developed for longitudinal brain age analysis (LSSL, LNE, LILAC, etc), the proposed method is novel in terms of local patch treatment and deformation-based pairwise learning. The sample size is acceptable. The results seem to be stronger than LILAC, the SOTA. These strengths outweighs the minor weaknesses.
