List of Papers Browse by Subject Areas Author List
Abstract
Longitudinal brain Magnetic Resonance Imaging (MRI) enables direct observation of within-subject structural change, helping disentangle aging-related trajectories from inter-individual anatomical variation. Recent self-supervised methods have shown that encoding such longitudinal change yields representations useful for brain age prediction and disease detection. These methods, however, learn directly from volumetric images rather than from brain network representations, overlooking inter-regional connectivity patterns. Graph-based approaches, on the other hand, model brain networks but typically at a single parcellation level, using cross-sectional data, without distinguishing normal aging from pathological change. We therefore propose HiLoGraph, a self-supervised framework that encodes multi-scale brain networks derived from morphological features and structural connectivity into a shared latent space, where longitudinal constraints enforce temporal consistency across within-subject visits. Trained exclusively on healthy subjects, the learned latent space serves as a normative reference from which subject level deviations can be quantified for downstream disease detection. Evaluated on 1,373 healthy subjects from the Aging Adult Brain Connectome (AABC) dataset and an independent clinical cohort of 156 subjects, HiLoGraph achieves an AUC of 0.769 for three-class classification of cognitively normal (CN), Alzheimer’s disease (AD), and Lewy body dementia (LBD), and 0.731 for AD vs. LBD, outperforming single-scale and non-longitudinal ablations. Code is available at https://github.com/tongchen2010/hilograph.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/3957_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to the Code Repository
https://github.com/tongchen2010/hilograph
Link to the Dataset(s)
N/A
BibTex
@InProceedings{CheTon_HiLoGraph_MICCAI2026,
author = { Chen, Tong AND Chen, Minheng AND Zhang, Jing AND Cao, Chao AND Su, Li AND Liu, Tianming AND Zhu, Dajiang},
title = { { HiLoGraph: Hierarchical-Longitudinal Brain Network Representation Learning } },
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
The paper proposes HiLoGraph, a self-supervised framework for learning representations from longitudinal brain MRI using multi-scale graph modeling. It combines hierarchical graph representations (ROI + fine-grained), longitudinal consistency across visits, and a normative modeling setup trained on healthy subjects, with deviations used for downstream CN/AD/LBD classification.
- 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 problem The idea of learning from healthy longitudinal data and using it for downstream disease detection is clinically meaningful. It avoids reliance on labels and fits well with the normative modeling literature in neuroimaging.
- Good combination of ideas While individual components are not new, combining multi-scale graph learning, longitudinal self-supervision, and normative modeling is interesting and works well in practice.
- Solid architecture + ablations The hierarchical design is sensible, and the clearly support the method.
- External validation Testing on an independent clinical dataset strengthens the paper.
- 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.
- Very small gains over strong baselines The improvements over the best baseline (LSSL) are quite small. For example, the 3-class AUC goes from 0.758 to 0.769, which is within the reported standard deviation. For AD vs LBD, performance is identical (0.731). It’s hard to tell if the added complexity provides a meaningful benefit.
- No statistical comparison with baselines The permutation test only shows performance above chance. There is no statistical comparison against the baselines, which seems important given how small the differences are.
- Limited evaluation / small clinical cohort The external dataset is relatively small (n=156, with only 20 CN), which makes the results harder to assess in terms of robustness. It would help to report variability (e. g. , across splits) or confidence intervals.
- Limited interpretability The model outputs deviation scores, but it is unclear which brain regions or graph components drive the predictions. Some form of regional or node-level analysis would make this more clinically useful.
- Clarity issues Some acronyms (e. g. , LBD) are not defined at first use.
- Figures are hard to interpret Fig. 2b is difficult to follow and would benefit from a clearer explanation. For Fig. 3, I do see some patterns that capture age structure for most methods. For example, why do you mention that it is not clear for fold 0 of 3HG-only?
- 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?
Overall, this is a well-motivated and solid paper. The method is well put together, the ablations are strong and the external validation is a clear plus. That said, the gains over strong baselines are relatively small, so it’s not entirely clear how much practical improvement the method brings. The lack of statistical comparison with baselines and the small external cohort also make the results a bit harder to assess. Despite these issues, the idea is very interesting so I lean slightly positive.
- Reviewer confidence
Somewhat confident (2)
- [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 #2
- Please describe the contribution of the paper
The paper proposes HiLoGraph, a self-supervised framework for learning hierarchical and longitudinal brain network representations from multimodal neuroimaging data.
The method combines: -multi-scale graph representations (ROI-level + fine-grained 3HG graphs), -a hierarchical graph encoder with cross-scale fusion longitudinal constraints enforcing temporal consistency across visits.
The model is trained exclusively on healthy subjects to learn a normative latent space, from which subject-level deviations are computed and used for downstream disease classification. The approach is evaluated on a longitudinal healthy cohort and an external clinical dataset, showing improved performance over single-scale and non-longitudinal baselines.
- 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 gap by jointly modeling: -multi-scale brain structure (ROI + fine-grained landmarks), -longitudinal dynamics This combination is novel and well justified, as prior work typically considers only one of these aspects.
2.The framework is well designed: -hierarchical GCN encoder with cross-scale fusion, -VAE latent space, -top-down decoder enforcing consistency. The design is coherent and aligns well with the problem formulation.
3.The use of temporal smoothness and contrastive constraints is a strong point. These losses explicitly enforce biologically meaningful trajectories and improve representation quality.
4.The paper includes: -multiple baselines (MLP, GCN, GAT, prior longitudinal methods), -ablations (ROI-only, 3HG-only, no longitudinal), -cross-validation. Table 1 shows consistent improvements of the full model over baselines.
5.Training only on healthy subjects and computing deviation scores is a principled and clinically relevant approach, consistent with recent literature.
- 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.A substantial part of the analysis (e.g., Fig. 2 and Fig. 3) relies on t-SNE visualizations to support claims about: -age structure in latent space, -separation between groups. These projections are not reliable for assessing global structure or distances. While classification results are provided, the interpretability claims about the latent space are largely qualitative.
2.The evaluation dataset has only 156 subjects, class imbalance (CN = 20 vs LBD = 80). This may affect the robustness of the reported AUC values.
3.Disease prediction is performed using logistic regression on deviation features. While this isolates representation quality, it also limits performance and does not fully explore the potential of the learned embeddings.
4.The paper compares against some longitudinal baselines, but does not include stronger recent non-graph approaches (e.g., transformer-based or diffusion-based longitudinal models). This limits the positioning of the method.
- 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
The paper presents a well-designed and technically sound framework that addresses an important problem in longitudinal neuroimaging representation learning. The combination of hierarchical graph modeling and longitudinal constraints is particularly compelling. However, the evaluation of the learned representations could be strengthened by reducing reliance on qualitative visualizations and providing more rigorous quantitative analysis of the latent space and normative modeling properties.
- 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 paper proposes a well-motivated and technically sound framework for hierarchical and longitudinal brain network representation learning. The combination of multi-scale graph modeling and longitudinal constraints is novel and supported by a solid experimental design, including ablations and comparisons with multiple baselines. The results show consistent improvements over alternatives, and the use of a normative modeling framework trained on healthy data is particularly relevant for clinical applications. The main limitations lie in the evaluation of the latent space and normative properties, which relies partly on qualitative visualizations, and in the relatively small clinical dataset used for downstream validation. However, these issues do not undermine the overall contribution. Overall, I find the paper to be a meaningful contribution that is slightly above the acceptance threshold.
- 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 HiLoGraph, a self-supervised VAE-based framework that learns brain network representations from multi-scale graphs (3HG landmarks + Destrieux atlas ROIs) constructed from morphological features and structural connectivity. The model is trained exclusively on 1,373 healthy AABC subjects with a longitudinal smoothness + contrastive loss, and the frozen encoder is then used to derive normative deviation scores on an external 156-subject clinical cohort (CN/AD/LBD). The full model reaches a 3-class AUC of 0.769.
- 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 correctly identifies a genuine gap: existing longitudinal SSL methods (LSOR, LNE, LSSL) operate on volumetric data, while graph-based methods are largely cross-sectional and trained on mixed cohorts. Combining hierarchical graphs, longitudinal constraints, and pure healthy training is a reasonable and clinically meaningful contribution. 2.Following Rutherford et al. for age/sex-conditioned deviation scoring on a frozen encoder, then testing on a held-out clinical cohort, is methodologically clean and avoids label leakage. 3.The four-way comparison (ROI-only, 3HG-only, w/o longitudinal, full) is informative, and Fig. 3 nicely shows that the age manifold structure is fragile without both hierarchy and longitudinal constraints. Table 2’s feature comparison is also a good touch and supports the deviation-feature framing.
- 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.AABC was acquired via the HCP-Aging protocol; the clinical cohort uses different scanners, sequences, and a separate preprocessing pipeline (FreeSurfer + FSL + DSI Studio vs. HCP minimal pipeline). No harmonization (e.g., ComBat, longitudinal ComBat, or a learned domain adaptation step) is reported. Apparent “deviations” in AD/LBD subjects may be partly confounded by site/scanner effects rather than true disease signal. 2.For comparison, LNE and LSSL. How were they adapted to graph inputs? Were they re-implemented from scratch? Trained on the same AABC subset with the same QC? Without these details, the baseline comparisons in Table 1 are not interpretable, and the claim that HiLoGraph beats them isn’t supportable. 3.”~400 nodes” is reported for 3HG parcellation, but 3HG extraction yields per-subject variable counts. How is the membership matrix M handled when N differs across subjects? How does the encoder handle variable-sized 3HG graphs in batches? 4.The paper produces 64-dim deviation vectors but never asks which dimensions (or which underlying ROIs/3HGs) drive the AD vs. LBD separation. Even a simple per-dimension AUC or a feature-importance plot would substantially strengthen the clinical narrative.
- 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
1.Please furthen review the paper. There are a few typos. p.5: “f_age isa two-layer MLP”, might be a typo; Fig. 2(c): KW p = 7.5e-04 is shown in the figure but reported as 4.1e-3 in text; The baseline methods, LSOR, LNE and LSSL just pops out without the full name.
- 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 a real gap with a reasonable architectural recipe, the experiments are organized sensibly, and the latent-space visualizations are convincing for the healthy training data. However, the weaknessnes mentioned about limit my enthusiasm.
- Reviewer confidence
Very confident (4)
- [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
Author Feedback
N/A
Meta-Review
Meta-review #1
- Your recommendation
Provisional Accept
- 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.
Reviewers are generally enthusiastic about the proposal and believe it addresses a really technical gap. The authors are encourage to clarify on the details raised by the reviewers and realistically discuss limitations of the method in a constructive way.