List of Papers Browse by Subject Areas Author List
Abstract
Functional MRI (fMRI)-derived dynamic brain functional connectivity (dBFC) holds significant potential for characterizing brain disorder-specific abnormalities in macroscopic neural activity patterns underlying cognition and behavior. Yet, standard approaches usually estimate connectivity via zero-lag correlations. This simplification ignores temporal lags among different brain regions, which are biologically meaningful markers of directed neural propagation and diverse pathological brain states. Furthermore, existing graph learning models often underexploit the brain‘s hierarchical modular architecture and struggle to capture disorder-induced community switching of brain regions between functional modules. In this work, we propose the Community-aware Dynamic Graph Neural Network (CaDGNN) framework to learn the spatiotemporal interaction patterns in lag-aware dBFC networks for brain disorder diagnosis. First, CaDGNN incorporates a lag-aware dBFC network construction module to encode spatiotemporal propagation patterns of brain activation by explicitly incorporating inter-regional time delays. Then, a temporal community co-occurrence encoder is introduced to extract temporal community structures and model dynamic inter-regional interactions. Based on this, a dual-level dynamic learning module models the switching behavior of graph nodes between communities and the temporal evolution of community structures simultaneously for final diagnosis. Evaluations on four independent datasets demonstrate that CaDGNN outperforms state-of-the-art methods. Our findings highlight the diagnostic value of incorporating both lag-aware dynamics and modular structure into graph learning. The codes of CaDGNN are available at https://github.com/Alyrrraa/CaDGNN.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/2999_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to the Code Repository
https://github.com/Alyrrraa/CaDGNN
Link to the Dataset(s)
N/A
BibTex
@InProceedings{LiQiu_CommunityAware_MICCAI2026,
author = { Li, Qiuyan AND Li, Guangming AND Yu, Wei AND Cheng, Guanxun AND Yu, Keyan AND Fan, Xiang AND Wang, Kai},
title = { { Community-Aware Dynamic Graph Learning for Brain Disorder Diagnosis } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 16886},
month = {September},
page = {pending}
}
Reviews
Review #1
- Please describe the contribution of the paper
This work proposes a Community-aware Dynamic Graph Neural Network which constructs lag-aware dynamic functional connectivity networks to explicitly model inter-regional time delays. It further integrates community structure into graph learning to capture modular brain organization. The proposed method is validated on multiple datasets for brain disorder diagnosis.
- 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 framework for brain disorder diagnosis by integrating lag-aware dynamic functional connectivity with community-aware graph learning. 2.The proposed dual-level modeling provides a multi-scale perspective for capturing both local and global brain dynamics. 3.The method is evaluated on multiple datasets, and includes basic interpretability analyses.
- 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 claimed novelty is limited. The overall framework is conceptually similar to dFCExperts(Chen, Tingting, et al. “dfcexpert: Learning dynamic functional connectivity patterns with modularity and state experts. “ IEEE transactions on medical imaging (2025). ), with differences mainly in specific implementations. In addition, lag-based functional connectivity is not new. The method appears as a combination of existing techniques rather than being driven by a unified and well-justified research motivation. 2.The concatenation of zero-lagged matrices in Fig. 1(d) lacks clear explanation and justification. 3.The performance on ADNI is unusually high, especially for the CN vs. EMCI task, and is significantly higher than competing methods, raising concerns about fairness and possible data leakage. 4.The performance of baseline methods is generally low. Important related methods, particularly those based on dynamic FC and community detection, are missing from the comparison. 5.Fig. 3(a) lacks explanation. 6.The methodology behind the Visualization Analysis in Section 3.3 is not clearly described. 7.In Fig. 2, grouping the same metrics together would improve readability and comparison.
- 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.
(2) Reject — should be rejected, independent of rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
Limited novelty, and the experimental results are difficult to interpret.
- 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 author appropriately addressed the concerns.
Review #2
- Please describe the contribution of the paper
This paper introduces CaDGNN, a community-aware dynamic graph neural network designed for brain disorder diagnosis using rs-fMRI data. The authors move beyond standard zero-lag connectivity by proposing a lag-aware network construction that explicitly models inter-regional time delays in neural propagation. To capture the brain’s evolving modular structure, they design a temporal community co-occurrence encoder. Finally, the method integrates a dual-level graph learning module to simultaneously capture intra-community node switching and inter-community temporal evolution.
- 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 shift from instantaneous (zero-lag) functional connectivity to lag-aware dynamic connectivity is well-motivated and better reflects the actual neurophysiological propagation of brain signals. 2.The experimental setup is very thorough. The authors validated their framework on four distinct datasets (ADNI, a local dataset, ABIDE, and ADHD-200) covering three different disorders. This demonstrates generalizability. 3.CaDGNN consistently outperforms a solid range of recent baselines (e.g., STA-GIN, BMR, BISTFormer) across multiple evaluation metrics, showing notable improvements in ACC and AUC.
- 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 overall lag-aware framework and hierarchical design are novel, the downstream learning modules heavily rely on off-the-shelf components, such as standard Transformer encoders and Graph Attention Networks (GAT). The algorithmic contribution lies more in how the data is pre-processed and structured rather than in the neural network operators themselves. 2.The lag-aware network construction relies on a three-point parabolic interpolation of the cross-covariance function to estimate delays. The authors do not adequately discuss or ablate how robust this heuristic is against the typical noise and artifacts found in rs-fMRI data. 3.The title of Section 2.2 contains a clear typo: “Grpah Learning” should be corrected to “Graph 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
It would be beneficial to include a brief discussion on the computational overhead introduced by the lag-aware dBFC construction and the multilayer network analysis, especially when compared to simpler zero-lag methods.
- 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 authors propose a biologically sound approach to graph learning for brain networks by incorporating temporal delays and dynamic community structures. The extensive validation across four datasets provides good evidence of the method’s effectiveness. However, the core neural network architectures lack significant novelty, as they largely rely on existing Transformer and GAT modules. Additionally, the lack of robustness analysis for the lag estimation and the missing details for the local dataset are concerns that limit the overall impact. If the authors can address the robustness of the pre-processing and provide more context for the local dataset in their rebuttal, it would strengthen the case for acceptance.
- 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 new framework for brain disorder diagnosis, called CaDGNN. Its main contribution has three parts. First, it introduces lag-aware dynamic functional connectivity into the model, instead of using only zero-lag correlations. Second, it adds temporal community detection to capture how brain regions change their community membership over time. Third, it uses a dual-level graph learning design to model both ROI-level dynamics and community-level temporal evolution. Overall, the paper aims to combine delay information, modular organization, and temporal dynamics in one diagnostic framework.
- 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 this paper is that its method is relatively complete. It does not only use static functional connectivity. It also does not rely on a simple temporal model. Instead, it combines lag information, community structure, and temporal learning in one framework. This gives the method some novelty. The experiments are also carried out on multiple datasets and tasks, including ADNI, ABIDE, ADHD-200, and a local MCI dataset. The results show that the method outperforms several baselines on some tasks. Another strength is that the paper includes an interpretation analysis of discriminative ROIs and connections, which makes the model more informative.
- 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 lag values are estimated at the individual level. However, the manuscript does not show whether the learned lag patterns are robust to inter-site heterogeneity. Since multiple datasets and different preprocessing pipelines are involved, the reported lag effects may be influenced by site-specific acquisition or preprocessing differences. 2 The merging of clusters across different time points relies on the assumption that the same community label is directly comparable over time. This assumption is important, but it is not sufficiently validated in the current manuscript. The authors should provide evidence for cross-temporal label consistency, examine the effect of the temporal coupling parameter ω, and clarify how the method avoids artificial merging caused by label switching. 3 The study uses the Schaefer-500 atlas, which is a very high-resolution parcellation. This choice increases feature dimensionality and model complexity, and may also increase the risk of overfitting. The manuscript does not show whether the performance gain is robust across different atlas resolutions. Repeating the experiments with lower-resolution parcellations, such as Schaefer-100 or Schaefer-200, would strengthen the methodological validity. 4 The community-level graph is constructed with an all-ones adjacency matrix, so each time point is fully connected to all others. This design weakens the explicit modeling of temporal order and may bias the model toward global similarity rather than local temporal evolution. The authors should evaluate more constrained temporal connections, such as local time windows neighborhoods, to justify the current design and clarify the contribution of temporal structure.
- 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 method combines lag-aware dynamic connectivity, temporal community modeling, and dual-level graph learning in a unified way. This gives the paper some methodological novelty. The experiments are also fairly extensive, with multiple datasets, baseline comparisons, ablation studies, and visualization results. Although several methodological issues remain, the paper is well organized and the overall results are promising.
- 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
We sincerely thank the reviewers for their constructive feedback. Below we address the concerns. Q1(R1-1, R2-1): Novelty. CaDGNN and dFCExperts follow fundamentally different paradigms. dFCExperts decouples spatial state from temporal state aggregation on zero-lag dBFC; CaDGNN jointly models spatiotemporal coupling in lag-aware dBFC networks. The key distinction is the CCE module, which couples the spatiotemporal features by tracking each node’s switching history across communities over time, a mechanism absent in dFCExperts. Ablation (Fig.2) confirms removing CCE causes larger performance drop than removing Transformer encoder, validating that improvements stem from spatiotemporal community modeling rather than standard operators. We acknowledge Transformer/GAT are established components; our contribution is their principled composition for spatiotemporal brain network modeling. Q2(R2-2, R3-1): Lag Robustness. Three-point parabolic interpolation mitigates the discrete TR sampling limitation for lag estimation. It refines the CCF peak using only its two neighbors, avoiding noise-prone broader CCF regions while requiring no extra hyperparameters. ADNI, ABIDE, and * are processed with the same DPARSF pipeline. Results on the single-site * dataset further indicate lag effects are not solely driven by multi-site acquisition. Following prior work, we discount delays >4s and convert to TR-specific shifts, reducing artifact-prone long delays while keeping lag estimation comparable across datasets. Q3(R3-2,4): Temporal Community. CaDGNN does not detect communities per segment and artificially merge them. Instead, Temporal Community Detection (TCD) treats time windows as layers and links the same node across adjacent layers via coupling parameter ω. Community assignments are inferred through a single joint optimization, inherently encouraging temporal consistency while allowing network-driven switching. This design reduces the risk of artificial merging due to label switching. For the community-level graph, the all-ones adjacency enables flexible cross-window information exchange; GAT then learns nonuniform temporal attention, so windows are not equally aggregated. Ablation (w/o CG, Fig.2) validates this design’s contribution. Q4(R1-3,4): Fairness & Baselines. Two types of baselines are included, spanning dynamic brain activity models (ST2GCN, BISTFormer) and modularity-prior GNNs (BMR, similar to dFCExperts in its decoupled modeling way). We use strict subject-wise 5-fold cross-validation, ensuring that all samples from the same subject are assigned to the same fold to prevent subject-level leakage across train/test splits. No demographic or clinical variables are used as input features. As shown in Table 1, most compared methods achieve relatively high performance on ADNI, especially for CN vs. EMCI, indicating intrinsic separability of the diagnostic tasks rather than an artifact of our method. Q5(R3-3): Atlas Resolution. Schaefer-500 is widely adopted in brain network studies. Its finer parcellation captures functional boundaries coarser atlases may miss, important for detecting subtle disorder alterations. CaDGNN’s consistent performance across four datasets with varying acquisition suggests results are not driven by atlas resolution alone. Moreover, TCD naturally groups ROIs into communities, providing built-in regularization against overfitting at higher resolutions. Q6(R1-2,5,6): Clarifications. For Fig.1(d), the concatenation combines complementary information: X_{fc}^t preserves current synchrony, H^t encodes lag-aware community dynamics. Fig.3(a) shows sensitivity of k (KNN neighbors) and w (window length), color denoting ACC. For Sec.3.3, Grad-CAM is applied to ROI-level embeddings from intra-level graph learning; importance scores are computed from class-score gradients and averaged across subjects. Q7(R1-7, R2-3): Paper Readability. We apologize for the ‘Grpah’ typo. We will regroup Fig.2 by metrics and correct the typos.
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 proposes lag-aware dynamic functional connectivity network construction and community-based temporal clustering. The rebuttal should address reviewers’ critical questions, including the novelty of the proposed method, robustness, and a clear explanation of the lag pattern.
- 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 addressed reviewers’ major questions, including the novelty of the proposed method, robustness, and a clear explanation of the lag pattern.
Meta-review #2
- After you have reviewed the rebuttal and updated reviews, please provide your recommendation based on all reviews and the authors’ rebuttal.
Reject
- Please justify your recommendation.
This is a borderline paper. While the rebuttal is responsive to many of the critiques, I remain unconvinced about the importance of this method. The experimental results focus almost exclusively on diagnostic prediction and do not evaluate the novelty or robustness of the learned brain networks. I am also suspicious of the ADNI results. The proposed method achieves substantially higher classification performance than the baseline methods and the current literature based on rs-fMRI. This points towards information leakage, perhaps in the selection of hyperparameters. Ultimately, this is yet another “brain network extraction” paper that uses classification as a proxy for impact, and I lean towards rejection.
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.
The reviewers found the paper methodologically relevant and empirically well supported, with validation across multiple datasets. After rebuttal, the main concerns on novelty, robustness, lag modeling, and experimental fairness were addressed, and the reviewers did not raise further major objections. Acceptance is recommended.
