List of Papers Browse by Subject Areas Author List
Abstract
Brain effective connectivity (EC) is key to understanding causal neural interactions and brain organization. However, learning EC from single-modal brain data, such as functional magnetic resonance imaging (fMRI) or electroencephalography (EEG), is limited by the inability to simultaneously capture sparse temporal and spatial information. This paper proposes a novel multimodal sparse generative flow network (MSGFlowNet), which integrates fMRI and EEG data through an attention-guided encoder and employs a multi-head self-attention sparse Transformer to extract features from the fused data. These features are then processed by two output heads of the generative flow network: one computes state transition probabilities and updates the mask, while the other determines the probability of generating a termination state. Experiments on synthetic and real-world datasets demonstrate that MSGFlowNet significantly outperforms state-of-the-art methods.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2025/paper/5411_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{SuZhi_MSGFlowNet_MICCAI2025,
author = { Su, Zhihao and Zhai, Jihao and Ji, Junzhong and Liu, Jinduo},
title = { { MSGFlowNet: Learning Effective Connectivity Network based on Sparse Generative Flow Network from fMRI and EEG Data } },
booktitle = {proceedings of Medical Image Computing and Computer Assisted Intervention -- MICCAI 2025},
year = {2025},
publisher = {Springer Nature Switzerland},
volume = {LNCS 15960},
month = {September},
}
Reviews
Review #1
- Please describe the contribution of the paper
The author developed a multimodal sparse generative flow network (MSGFlowNet), which integrates fMRI and EEG data to identify effective connectivity. The model’s performance has been tested on both simulated datasets and real-world fMRI and EEG data.
- 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 manuscript is well-composed, with only minor typographical errors. The proposed framework was validated using both simulated and real-world data, demonstrating positive outcomes.
- 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.
Please see the comments below.
- 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.
- 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
Major comments:
- Time Lag Information and Method Comparison
In Section 1, the authors mention that RNNs are suitable for identifying time lags in Granger causality analysis. However, the paper does not appear to utilize time lag information in the subsequent methodology. Could the authors please comment on this discrepancy? Additionally, it would be beneficial to compare the proposed method with prior approaches that explicitly consider time-varying and nonlinear Granger causality, such as:
- Chuang, Kai-Cheng, et al. “Nonlinear conditional time-varying Granger causality of task fMRI via deep stacking networks and adaptive convolutional kernels.” MICCAI, Springer, 2022.
- Li, F., et al. “Unified model selection approach based on minimum description length principle in Granger causality analysis.” IEEE Access, 8, 68400–68416 (2020).
- Data Concatenation and Notation Consistency
In Section 3.2 and Figure 1, please clarify how the fMRI and EEG data are concatenated. The labels appear inconsistent between Equations (1)–(3) and Figure 1(1). Specifically:
- Where is matrix A represented in Figure 1?
- Is the fMRI data represented as Nf×Tf or Lf×Tf? Please clarify.
- Are both H and I trainable parameters?
- Additionally, why was the model trained on source-level time series rather than on the measured time series?
- Spatial Resolution and ROI Definition In Section 4.2, please provide the spatial resolution for both the fMRI and EEG data. Regarding ROIs, was a single voxel used for each ROI, or was the signal averaged across multiple voxels?
- Classification Model Specification In Section 4.2, please specify which classification model was used for the classification task.
Minor comments:
- Typographical Error in Section 1 There is a typographical error in Section 1: “deep learning methodsv [18]” should be corrected to “deep learning methods [18]”.
- Justification of Effective Connectivity Findings Rather than relying solely on classification task performance to validate the identified effective connectivity in real-world data, please consider providing additional justification. For instance, referencing prior clinical findings or established neurophysiological relationships among the implicated brain regions would help support the plausibility and interpretability of the connectivity patterns discovered by the model.
- Time Lag Information and Method Comparison
In Section 1, the authors mention that RNNs are suitable for identifying time lags in Granger causality analysis. However, the paper does not appear to utilize time lag information in the subsequent methodology. Could the authors please comment on this discrepancy? Additionally, it would be beneficial to compare the proposed method with prior approaches that explicitly consider time-varying and nonlinear Granger causality, such as:
- 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 — could be rejected, dependent on rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
The methodological descriptions lack sufficient clarity, making it difficult to fully assess the validity of the proposed approach, despite the reported positive model performance.
- 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 addressed my major comments. The updates and revisions can be feasibly implemented in the final version of the paper.
Review #2
- Please describe the contribution of the paper
The paper presents an innovative approach that leverages multimodal neuroimaging data and a novel sparse generative framework to improve causal brain connectivity modeling.
- 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 model effectively fuses fMRI and EEG data using an attention-guided encoder, overcoming the limitations of single-modal data in capturing both temporal and spatial dynamics, including the following particular architectures: 1). Sparse Transformer architecture: A multi-head self-attention sparse Transformer is used to extract features from the fused multimodal signals, enhancing the model’s ability to represent complex neural interactions. 2). Dual-head generative flow network design: the model includes two output heads: one estimates state transition probabilities and updates masks, while the other determines the termination state probability, allowing for better modeling of causal structures.
- 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.
Major issue:
Testing the algorithms on synthetic data (as done in the first experiment) is largely uninformative. I do not doubt the capacity of neural networks to integrate multimodal signals. However, the real challenge in such problems lies in handling high SNR data, which synthetic data often fails to capture. Therefore, it is essential to evaluate your method on real-world datasets and clearly document any necessary preprocessing steps. Furthermore, the results in Table 1 offer little insight: the performance differences between methods are minimal, and with some tuning, your baselines could likely perform significantly better.
Minor issues:
-
The paper spends a considerable amount of space reintroducing well-established theoretical methods that have already been published in major journals. Since this work is positioned as an application-focused study, you should streamline the methodological background and instead highlight how these techniques are adapted to your specific problem.
-
The manuscript frequently uses strong terms like “significant improvement” and “excellent results” to describe the performance, yet the experiments do not convincingly support these claims.
-
Figure 3 is rather confusing. Why are different methods connected with dashed lines? This visualization lacks clarity and may mislead the reader.
-
- 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.
- 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?
Interesting problem and novel tools, but the experimental design is incomplete and flawed.
- 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 presents MSGFlowNet, a novel method for learning effective brain connectivity (EC) networks from multimodal neuroimaging data, specifically fMRI and EEG. The key contribution lies in redefining EC estimation as a generative task using a sparse generative flow network (GFN) framework. To address the complementary characteristics and limitations of fMRI (high spatial resolution) and EEG (high temporal resolution), the authors introduce an attention-guided multimodal encoder to align and fuse both data sources. A sparse Transformer is used to extract key features from the resulting high-dimensional sparse fused data. Finally, the fused features are passed into the GFN, which estimates causal brain connectivity by modeling the probability distributions of state transitions and terminations. The approach is evaluated on both simulated datasets—demonstrating strong performance across metrics like precision, recall, F1, and SHD—and a real-world simultaneous EEG-fMRI dataset from OpenNeuro for brain state classification (rest vs. sleep). In all experiments, MSGFlowNet outperforms state-of-the-art EC learning methods, particularly in terms of precision and recall, validating its effectiveness in both synthetic and real data settings. The code is also made publicly available, supporting reproducibility.
- 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 introduces a first-of-its-kind generative flow-based model for effective connectivity estimation using both fMRI and EEG, presenting a clear methodological advancement over existing approaches. The multi-stage pipeline—consisting of multimodal fusion, sparse feature extraction, and EC graph generation—is thoughtfully constructed, and each component is backed by solid theoretical rationale. The use of a sparse Transformer is especially well-suited to handle the inherent high-dimensional sparsity resulting from fMRI-EEG fusion, allowing the model to scale efficiently while maintaining interpretability and performance. Furthermore, the generative flow network formulation of EC as a probabilistic graph generation problem is novel in this domain and aligns well with causal graph modeling principles. The evaluation strategy is comprehensive, using both simulated datasets (which help assess performance in controlled settings) and real EEG-fMRI data (which provides real-world relevance).
- 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.
While the paper is strong overall, there are following areas that could benefit from further clarity or expansion.
- limited interpretability analysis of the learned EC graphs—while classification performance is high, there is no qualitative or neuroscientific discussion on which brain regions or connections are most influential, especially for distinguishing rest and sleep states. This would enhance the clinical relevance of the method.
- The computational overhead of using generative flow networks is acknowledged but not evaluated—training time, memory usage, and inference speed comparisons with other models would be valuable, especially for large-scale applications.
- Some of the parameters like (Lf, Le, I, H) used in equations (1), (2) are important for robustness analysis, which are not clearly discussed.
- In section 4.1, why only 5 brain regions are selected from simulated fMRi data ?
- The real dataset used is relatively small (248 samples in total); although the model performs well, future versions would benefit from tests on larger and more diverse datasets to confirm generalizability.
- 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.
- 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?
I recommend this paper for weak acceptance due to its good methodological innovation and consistent performance improvements over the state of the art. MSGFlowNet introduces a novel and effective strategy for learning brain effective connectivity from multimodal neuroimaging data using a generative flow-based approach. The model is thoughtfully designed, integrates fMRI and EEG features in a principled way, and addresses the high-dimensional sparsity that plagues many multimodal methods. While the paper could improve in terms of interpretability, scalability analysis, and broader comparisons, its strengths in performance, originality, and relevance to the MICCAI community clearly justify its inclusion.
- 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.
Author almost addressed my comments.
Author Feedback
We thank all reviewers for their valuable and professional comments and suggestions. All reviewers recognized the novelty of our method, clarity (well structured &written), comprehensive evaluation, open code and also point out some weaknesses to help us improve our paper. We will address the concerns raised by 3 reviewers. (below Ax means answer to x-th Major concerns, Ix means answer to x-th Minor concerns).
Public responses: To Reviewer#2 (A2), Reviewer#3 (A3): We apologize for the confusion. Lf and Le denote fMRI regions and EEG channels, unified as Nf and Ne to match Figure 1. Matrix A will be added to Figure 1 for clarity. Both H and I are trainable, implemented as learnable linear layers and optimized end-to-end. Training on source-level time series better captures neural activity by solving modality-specific inverse problems. To Reviewer#2 (I2), Reviewer#3 (A1): Effective connectivity(EC) analyses showed increased links between the right supramarginal gyrus and left precuneus during rest versus sleep. These regions support attention, self-awareness, and sensory integration. The supramarginal gyrus processes sensory input and spatial attention, while the precuneus relates to self-referential thought. Enhanced connectivity reflects cognitive readiness and environmental monitoring in rest, reduced during sleep with lower consciousness.
To Reviewer#1: A1&I2:1)While synthetic data has limitations in modeling real fMRI and EEG characteristics, we follow prior works [14][28] to ensure comparability and controlled evaluation. 2)To address this, we also validated our method on simultaneous fMRI-EEG sleep-rest data. Preprocessing was performed using SPM12 and EEGLAB following [8]. 3)Thank you for your valuable suggestion. In the revised manuscript, we have adopted a more cautious wording to describe the performance improvements and avoid overly strong claims. Paired t-tests on F1 scores indicate that MSGFlowNet shows statistically significant advantages over Two-Step (p=0.01), MetaRLEC (p=0.01), and DiffAN (p=0.04). I1:We revised the Methods section to briefly introduce generative flow networks and place more emphasis on how fMRI and EEG source-level time series are combined. We also clarified that the Sparse Transformer is used to extract important features. I3:We apologize for the confusion caused by Figure 3.To avoid confusion caused by the dashed lines in Figure 3, we have replaced it with a clearer bar chart for more direct method comparison.
To Reviewer#2: A1:While RNNs model time lags explicitly, MSGFlowNet captures temporal dependencies implicitly via neural ODE-based flow and message passing for dynamic cross-modal interactions. Due to committee constraints, we cannot add experiments but will revise the introduction to discuss and cite related methods. A3:The fMRI data had a spatial resolution of 6 mm³, and signals were averaged across voxels within each ROI. EEG data were downsampled to 250 Hz. A4: In Section 4.2, we used an SVM with RBF kernel to classify rest and sleep states from EC graphs. I1:We acknowledge the typo in “methodsv [18]” and appreciate the reviewer’s attention to detail.
To Reviewer#3: A2:Due to space limits, we did not include computational overhead analysis here.In our experiments, training MSGFlowNet on a single subject’s real data takes 22 minutes on an NVIDIA RTX 4090 GPU. A4:The simulation setup follows the design in [14][28], which uses 5 brain regions for evaluating EC methods. Therefore, we set it to 5 to align with current mainstream studies. Following your suggestion, we will expand the experiments to include more regions. A5:We agree a larger dataset would improve generalizability. Due to cost and strict requirements, the 248 samples used are the largest publicly available simultaneous fMRI-EEG dataset.
We hope our response addresses 3 reviewers’ concerns and encourage a score increase to support our work.
Best regards, Authors of 5411
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”.
N/A
- 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