Abstract

The acquisition of structural brain network data is inherently challenging due to high costs of Diffusion Tensor Imaging (DTI) and the complexity of data processing such as tractography. Moreover, medical datasets often exhibit severe class imbalance where the sample size of healthy subjects highly exceeds that of diseased. While recent graph generation models offer a potential solution, its application to brain networks is understudied as they often underestimate preserving topological feature which is an essential biomarker. To address these limitations, we propose a conditional graph diffusion model that ensures high-fidelity graph generation by leveraging persistent homology. Specifically, we introduce a Conditional Graph Diffusion (ConGD) method that utilizes Condition Infused Attention (CIA) module with class conditioning, to enable the targeted synthesis of brain networks, and Topology Aligning (TA) regularization to enforce topological consistency. Experiments on the Alzheimer’s Disease Neuroimaging Initiative (ADNI) dataset demonstrate that our approach provides high-fidelity synthetic brain networks under label conditions, which are further validated for improving predictive performance through downstream graph classification tasks.

Links to Paper and Supplementary Materials

Main Paper (Open Access Version): https://papers.miccai.org/miccai-2025/paper/4024_paper.pdf

SharedIt Link: Not yet available

SpringerLink (DOI): Not yet available

Supplementary Material: Not Submitted

Link to the Code Repository

N/A

Link to the Dataset(s)

N/A

BibTex

@InProceedings{ParJoo_Conditional_MICCAI2025,
        author = { Park, Joonhyuk and Lee, Donghyun and Wu, Guorong and Kim, Won Hwa},
        title = { { Conditional Graph Diffusion with Topological Constraints for Brain Network Generation } },
        booktitle = {proceedings of Medical Image Computing and Computer Assisted Intervention -- MICCAI 2025},
        year = {2025},
        publisher = {Springer Nature Switzerland},
        volume = {LNCS 15971},
        month = {September},
        page = {227 -- 236}
}


Reviews

Review #1

  • Please describe the contribution of the paper

    The paper introduces ConGD, a conditional graph diffusion model for generating structural brain networks with preserved topological features. It combines class-aware attention and topological regularization, leading to more realistic synthetic graphs and improved classification on an Alzheimer dataset.

  • 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 proposes ConGD, a graph diffusion model that preserves global topological structure using persistent homology and a Topology Alignment regularizer. This is important for brain networks, where topological features often capture meaningful clinical patterns that standard models overlook.

    2. Their Condition Infused Attention module injects class and structural cues into the diffusion process, guiding the model to generate graphs that match both label and structure. This helps address class imbalance and improves the realism of class-specific brain networks, compared other general generative models.

    3. The authors demonstrate the clinical relevance of ConGD by applying it to the ADNI dataset, a well-known benchmark in Alzheimer’s research. The generated graphs are evaluated not just for realism but for their utility in downstream classification tasks, showing that they enhance predictive performance.

  • 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 method is only tested on the ADNI dataset for Alzheimer’s, so it’s hard to tell how well it would work for other brain disorders. Since brain network patterns can differ a lot across conditions like autism (ABIDE), the results may not generalize. More diverse evaluation would strengthen the case for clinical use.

    2. While the authors compare their model to recent strong generative models, they doesn’t compare against topology-aware models like TopoGAN or TopoDiff. Without these baselines, it’s unclear how much the proposed Topology Alignment module improves over prior work.

    3. The paper lacks an ablation study to separately assess the contributions of the Topology Alignment module and the Condition Infused Attention module. Without this, it’s unclear how much each component individually contributes to the model’s performance and whether one is more critical than the other.

    A minor comment: Persistence landscape is an older vectorization method for persistent homology. Using persistence images or silhouettes with the appropriate power would likely enhance the topological module.

  • 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 does not mention open access to source code or data but provides a clear and detailed description of the algorithm to ensure reproducibility.

  • 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/2025/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 — could be 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 presents a novel approach to graph generation with significant potential for neuroimaging applications, particularly in preserving topological features. However, the lack of a broader evaluation across datasets and an ablation study makes it difficult to fully assess the individual contributions of the proposed components.

  • 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.

    I went over the rebuttal and other reviewers’ comments. The paper has solid contribution, and it should be accepted.



Review #2

  • Please describe the contribution of the paper

    The paper presented a conditional graph difussion model with the integration of the condition infused attention and topology aligning for graph data embedding, and applied to the brain network reconstruction using the DTI imaging acquired from the ADNI dataset. For the downstream task, the ConDG serves as a data augmentation technique to handle the imbalance problem in brain image classification task.

  • 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 major strength of the paper lies in the integration of persistent homology techniques, including the persistence diagram and persistence landscape algorithms, to maintain consistency between the graph generated by the diffusion model and the original DWI-derived graph. In addition, the sampling method controls the distribution of the generated graphs, which is important for AD classification and ensures alignment within the same feature space. As a result, using the augmented graphs leads to improved performance.

  • 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 major weakness of this paper lies in the lack of clarity regarding implementation details and the effectiveness of the individual components in the proposed ConDG framework.

    1. Why is the model tested on DTI images? The motivation for using DTI for brain imaging classification is unclear. Can the proposed method also be applied to T1-weighted MRI or other modalities available in the ADNI dataset?

    2. What subjects were enrolled? To help readers reproduce the results, it is highly recommended to provide detailed subject inclusion/exclusion criteria.

    3. What types of features were included? The authors mentioned that node features are denoted by X and edge features by A, but no detailed description of these features was provided.

    4. The ConDG model includes the CIA and the class-conditioned topological embedding modules. However, no ablation study was conducted, making it impossible to assess the individual contribution of each module.

    5. Besides references [11] and [22], the authors should compare their method with state-of-the-art (SOTA) approaches in this domain. The compared methods are not tailored for brain network generation using DTI. Two potentially relevant studies are:

    [Zong, Y., Zuo, Q., Ng, M.K.P., Lei, B., and Wang, S., 2024. A new brain network construction paradigm for brain disorder via diffusion-based graph contrastive learning. IEEE Transactions on Pattern Analysis and Machine Intelligence.]

    [This reference is a duplicate of the one above. Consider replacing it with a different one if intended.]

  • 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 does not provide sufficient information for reproducibility.

  • 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/2025/en/REVIEWER-GUIDELINES.html

    There are some typos or uncleared statements in the presented manuscript, suas in section 3.1, ‘while the edge embedding A_k^t is derived from A_k^t’. Does this indicate a copying operation is performed?

  • 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 — could be accepted, dependent on rebuttal

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

    There is a lack of implementation details, subject enrollment criteria, and an ablation study. In addition, no comparison was conducted with the SOTA methods in this field.

  • Reviewer confidence

    Somewhat confident (2)

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

    Reject

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

    The motivation is still not clear and the implementation detials are still not provided.



Review #3

  • Please describe the contribution of the paper

    Authors propose a conditional graph diffusion model with topology aligning regularization. They validate their method on the ADNI dataset.

  • 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.

    Generating synthetic brain networks for disease conditions affecting topology is crucial for augmenting limited data to aid downstream analysis. The problem is well set up and the method is clearly explained.

  • 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.

    Experiment section is weak. Ablation study is not done to show the contribution of the CIA module. Figures or discussion showing saliency map and topological features being learnt by the proposed method would help concrete the findings.

  • 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 does not provide sufficient information for reproducibility.

  • 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/2025/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 paper is well written. The authors should consider providing github link for reproducing the results.

  • 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 addressed my concern about the ablation study.




Author Feedback

We thank every reviewer for their constructive feedback and effort given to review our paper. As the length of the rebuttal is limited, we will first address two common concerns raised by all reviewers, including the Meta-Reviewer.

1) Lack of comparison against topology-aware models and SOTA graph generation methods A) We are aware of existing topology-aware models for generative tasks. However, most are designed for domains such as images [1] or other non-graph data [2-3], making them inapplicable directly on graphs. Thus, the choice of baselines focused on diffusion-based graph generation methods, specifically GDSS and DiGress, as they are widely adopted and align with our approach. We appreciate the reviewers’ suggestion to include additional SOTA methods. For this submission, we prioritized a clear focus on diffusion-based models to maintain a well-defined experimental scope. Nonetheless, we agree on the importance of broader comparisons and will consider including them in a future journal extension.

2) Lack of ablation study A) We fully agree with the need for an ablation study on the CIA module of ConGD. We have, in fact, conducted an internal assessment of the CIA module’s impact, which we could not include in the manuscript due to space limitations, showing (average %p gain of acc, prec, rec, f1): Binary (+2.48, +1.93, +2.48, +2.60); 5-way (+4.43, +4.63, +4.43, +4.78). Notably, the gains are more pronounced in the challenging 5-way task, demonstrating the robustness of the CIA module. These improvements are remarkable given that ConGD’s average margin over the second-best results is (1.34, 2.74, 2.94, 4.10). To further validate ConGD, the ablation results on our regularizer R_TA are: Binary (+2.58, +2.48, +2.58, +2.63) and 5-way (+2.85, +2.45, +2.85, +2.40).

Below are the rebuttal for other concerns, with abbreviated terms: Meta-Reviewer (MR), Reviewer# (R#), Weakness# (W#).

MR-W1) Related work section did not cover multiple key works. A) We appreciate the reviewer’s concern regarding the lack of coverage of related works. We acknowledge other relevant works on the use of persistent homology in the medical field [4] and other diffusion graph generation methods [5]. After carefully reviewing their relevance, we will include a more comprehensive discussion of these works in the final version of the paper.

R1-W1) Only tested on ADNI dataset R3-W1) Why tested on DTI images? A) We selected the DTI ADNI dataset due to the availability of a preprocessing pipeline for obtaining structural brain networks, where the topological properties of graphs become critical. Therefore, we would like to emphasize that our method is not limited to the DTI ADNI dataset, but can be generally applied to any structural brain network. Given this generality, we expect similar performance improvements when applied to other structural brain network datasets, regardless of the specific neurological disorder. We agree with Rev 1 that extending our method to different brain disorder datasets may further demonstrate its generalization and practicality, which we will consider in a potential journal extension.

R3-W) Lack of clarity in implementation details A) Due to page limitations and the absence of supplementary materials, we could not include them in the manuscript. However, to ensure reproducibility, we will share our code upon acceptance.

[1] Wang, F., et al. “Topogan: A topology-aware generative adversarial network.” ECCV, 2020. [2] Zhang, Y., et al. “TopoDiff: Improving Protein Backbone Generation with Topology-aware Latent Encoding.” bioRxiv, 2023. [3] Hu, J., et al. “Topology-aware latent diffusion for 3d shape generation.” arXiv:2401.17603, 2024. [4] Bukkuri, A., et al. “Applications of topological data analysis in oncology.” Front. Artif. Intell., 2021. [5] Kong, L., et al. “Autoregressive diffusion model for graph generation.” ICML, 2023.




Meta-Review

Meta-review #1

  • Your recommendation

    Invite for Rebuttal

  • If your recommendation is “Provisional Reject”, then summarize the factors that went into this decision. In case you deviate from the reviewers’ recommendations, explain in detail the reasons why. You do not need to provide a justification for a recommendation of “Provisional Accept” or “Invite for Rebuttal”.

    The reviewers all acknowledge the methodological contribution. However, a few concerns remain, specifically the experimentation where the comparison methods may not represent the state of the art, e.g. topology-aware models like TopoGAN or TopoDiff or also previous graph generation approaches presented at MICCAI. Furthermore, the related works section appears to be missing multiple key works from Persistent homology applied to medical images as well as diffusion based graph generation.

  • 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. You may optionally write justifications for ‘accepts’, but are expected to write a justification for ‘rejects’

    N/A



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. You may optionally write justifications for ‘accepts’, but are expected to write a justification for ‘rejects’

    N/A



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. You may optionally write justifications for ‘accepts’, but are expected to write a justification for ‘rejects’

    N/A



back to top