List of Papers Browse by Subject Areas Author List
Abstract
Brain connectome provides a comprehensive map of neural connections within the brain, which is fundamental to understanding brain function and dysfunction. Recently, transformer-based methods have significantly advanced the analysis of brain connectome. However, these methods mainly rely on learning node-level representations, overlooking the critical role of brain connectivity (edge) to elucidate the mechanism of brain disorders, which limits their interpretability. Besides, the brain connectome exhibits a modular and hierarchical topology, while traditional Euclidean-based methods struggle to model these hierarchical structures without high distortion, limiting their representational capacity. To address these limitations, we propose BrainHTF, a novel hyperbolic manifold-based Transformer framework tailored for learning causal graph representation of brain connectome. Specifically, we first introduce a mask generator to disentangle the brain connectome into causal and bias subgraphs. Then, a hyperbolic transformer is utilized to learn their respective disentangled representations in hyperbolic space. Finally, causal intervention is conducted on representations for discovering invariant causal patterns for prediction. Experimental results on the ABIDE and ADNI datasets demonstrate that our BrainHTF consistently outperforms several state-of-the-art methods in brain disease detection while also identifying interpretable biomarkers. The code is available at https://github.com/MYzzZ123/BrainHTF.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/2442_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to the Code Repository
https://github.com/MYzzZ123/BrainHTF
Link to the Dataset(s)
ABIDE dataset: http://fcon_1000.projects.nitrc.org/indi/abide/
ADNI dataset: http://adni.loni.usc.edu/
BibTex
@InProceedings{SunQiy_BrainHTF_MICCAI2026,
author = { Sun, Qiyu AND Mao, Zhenlin AND Feng, Ruiqing AND Huang, Jiashuang AND Wang, Mingliang},
title = { { BrainHTF: Learning Causal Graph Representation of Brain Connectome with Hyperbolic Transformer } },
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 a hyperbolic Transformer framework for learning causal graph representations of brain connectomes. The core idea is to explicitly model brain connectivity by disentangling the input graph into causal and bias subgraphs via a learnable mask generator, and then embedding these subgraphs into a hyperbolic space to better capture the hierarchical structure of brain networks. A hyperbolic Transformer is designed to learn representations on the Poincaré ball manifold, followed by a causal intervention mechanism that promotes invariant causal feature learning by perturbing bias representations. The framework jointly optimizes classification, uniformity, and intervention losses. Experiments on two benchmarks demonstrate improved performance over GNN-, Transformer-, and hyperbolic-based 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 integration of causality and hyperbolic geometry is conceptually well motivated for brain connectome analysis. 2.Both empirical performance and interpretability have been demonstrated to show the effectiveness of the proposed method.
- 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.Only two datasets have been used for validation. More brain network datasets should be included to further demonstrate the effectiveness of the method across different brain diseases. 2.Brain network datasets usually contain limited training samples. It is not clear whether the data can support the training of the complex hyperbolic Transformer architecture. 3.The review of existing literature is limited. More recent related works should be discussed.
- 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
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?
See weaknesses
- 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
Review #2
- Please describe the contribution of the paper
The paper introduces BrainHTF, a novel framework designed to learn causal graph representations of brain connectomes. The method has three major components: (i) a subgraph generator that disentangles connectomes into causal and bias subgraphs, (ii) a hyperbolic transformer for modeling hierarchical graph structures, and (iii) causal intervention to identify invariant, interpretable patterns associated with brain disorders. Experiments conducted on the ABIDE and ADNI datasets show that BrainHTF outperforms existing methods, including GNN-based approaches. Further analysis highlights its ability to identify significant biomarkers.
- 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 paper combines causal–bias disentanglement, hyperbolic modeling, and causal intervention into a coherent framework. Each component targets a key limitation, and together they enable more robust and meaningful graph representations.
- The method improves representation capacity via hyperbolic geometry while explicitly modeling causal structure for interpretability. This allows both strong predictive performance and identification of meaningful biomarkers.
- The model outperforms a wide range of baselines, including GNNs, transformers, and hyperbolic methods.
- 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 mask generator operates at the level of individual edges, implicitly assuming that causal structure can be captured through pairwise interactions. However, real-world brain processes often involve multi-hop and higher-order dependencies, as well as region-level effects spanning groups of neighboring ROIs, which cannot be fully represented by one-step edge-level importance alone.
- The method identifies “causal” edges through a learned mask, but its stability across different folds or runs is not evaluated. This is important, since claims about causal patterns or biomarkers require some level of consistency; without it, it is difficult to tell whether the identified edges are truly meaningful or just artifacts of training variability.
- 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.
(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 overall proposes a solid methodological contribution, but some key aspects would benefit from deeper analysis, particularly the lack of thorough investigation of the causal aspect. Additionally, while the intuition behind the mask generator is in the right direction, its design remains somewhat simplistic.
- 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
This paper proposes a hyperbolic transformer model called BrainHTF. It is designed for classifying brain disorders from fMRI-derived brain connectomes by combining three ideas: splitting each connectome into a learned causal subgraph and bias subgraph, encoding those subgraphs with a hyperbolic transformer to better capture the hierarchical structure, and applying a causal intervention objective to encourage disease-relevant patterns that remain stable despite nuisance variation. The method is evaluated on two public datasets that show better classification performance than several GNN, transformers, and hyperbolic-graph baselines. The paper also argues for improved interpretability by highlighting disease-related regions and connections.
- 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 presents a clear motivation from limitations of prior work: weak edge-level interpretability and the difficulty of modeling hierarchical brain topology in Euclidean space. It has a clear idea of combining causal disentanglement and hyperbolic representation learning.
2.The empirical evaluation is reasonably convincing. It is supported by ablation studies showing that removing the hyperbolic component, mask disentanglement, or causal intervention reduces performance.
3.The paper is well written with clear overall diagram in Figure 1 and the visualization analysis in Figure 2.
- 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.Although the design is inspired by causality, what is presented in the paper is not strict biological causality. The input graph is built from Pearson correlations of fMRI signals, which are associative. Also, the intervention happens on learned latent representations, not on the brain or on an explicit structural causal model. So the claim of causality might be too strong in the paper.
2.Although the ablation studies show that replacing the hyperbolic transformer with a standard transformer reduces accuracy, the work does not thoroughly analyze when hyperbolic space is beneficial. Also, it is unclear why the curvature of the hyperbolic space is set to -3.It is also unclear how other hyperparameters are chosen.
3.Although this work proposes a new method for brain connectome analysis, the contribution is incremental and the components are not new. Hyperbolic transformers already existed in Hypformer, and causal-subgraph / intervention-based invariant graph learning was already established in graph machine learning.
- 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.
(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?
Although the paper mainly combines existing methods and ideas, this application to brain connectome analysis is still interesting and can stimulate discussion.
- 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
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.
After reviewing the three reviews (one Weak Reject, two Weak Accept), I find that all reviewers agree the paper has clear strengths. The idea of combining causal disentanglement and hyperbolic geometry for brain connectome analysis is well motivated. The empirical results are convincing, with ablation studies supporting the design choices, and the paper is well written. However, several concerns prevent a stronger recommendation.
The main issue is that the paper overclaims causality. The input graphs are built from Pearson correlations, which show associations, not true causes. The intervention happens on learned representations, not on the brain or a real causal model. The authors should soften their language.
Reviewer 2 raises two important points. First, the mask generator only looks at individual edges, but real brain processes often involve complex patterns across multiple regions. Second, the stability of the identified causal edges across different training runs is not evaluated, so it is unclear whether the reported biomarkers are reliable.
Reviewer 1 notes that only two datasets are used, and it is unclear whether the complex model can be trained effectively on the small sample sizes typical of brain network datasets. Reviewer 3 adds that the choice of hyperbolic curvature is not justified.
Regarding novelty, the individual components are not new – hyperbolic Transformers and causal intervention learning already exist. However, combining them for brain connectome analysis is still interesting and can stimulate discussion. This is why I lean toward Weak Accept rather than Reject.
In the final version of the paper, the authors should do three things: tone down the causality claims, evaluate the stability of the learned masks across multiple runs, and justify the curvature choice. Addressing these would strengthen the paper significantly.