List of Papers Browse by Subject Areas Author List
Abstract
Dynamic functional connectivity (dFC) derived from fMRI captures the temporal dynamics of brain networks, where cross-frequency features provide complementary characterizations for brain disorder classification. Although existing multi-band approaches incorporate sub-band decomposition, they primarily rely on simplistic averaging or fixed-weight strategies, failing to adaptively fuse information across multiple frequency bands. To handle this limitation, we propose a dual-stream multi-band fusion network (DSMFN): 1) The frequency-domain stream employs a sub-band graph encoding-interaction module, where local graph convolution networks (GCNs) extract band-specific topological features, and lightweight convolutions replace computationally intensive attention mechanisms for data-driven band contribution allocation, followed by a global GCN to aggregate cross-band information; 2) The time-domain stream preserves local dynamic properties via residual multi-layer perceptron networks; 3) A feature-temporal dual-dimension cross-attention mechanism jointly models temporal evolution and cross-domain complementarity to adaptively integrate multi-band features with time-varying characteristics. Experiments on two distinct brain disease datasets demonstrate the effectiveness of DSMFN, achieving accuracies of 91.40% for MCI and 70.18% for ASD classification. This study provides an efficient fusion framework for multi-band dynamic brain network analysis, advancing precise diagnosis of brain disorders. Our code is available at https://github.com/WuLingBNU/DSMFN.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2025/paper/1654_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to the Code Repository
https://github.com/WuLingBNU/DSMFN
Link to the Dataset(s)
ADNI-2 dataset: https://adni.loni.usc.edu/
ABIDE-II dataset: https://fcon_1000.projects.nitrc.org/
BibTex
@InProceedings{WuLin_DualStream_MICCAI2025,
author = { Wu, Ling and Li, Hexi and Lyu, Zhengyuan and Song, Zhiwei and Yu, Hu and Guo, Xiaojuan},
title = { { Dual-Stream Multi-Band Fusion Network for Dynamic Functional Connectivity Analysis in Brain Disorder Classification } },
booktitle = {proceedings of Medical Image Computing and Computer Assisted Intervention -- MICCAI 2025},
year = {2025},
publisher = {Springer Nature Switzerland},
volume = {LNCS 15971},
month = {September},
page = {289 -- 298}
}
Reviews
Review #1
- Please describe the contribution of the paper
The paper proposes a novel Dual-Stream Multi-Band Fusion Network (DSMFN) designed to enhance dynamic functional connectivity analysis for brain disorder classification. The framework integrates two complementary processing streams: a frequency-domain stream that employs a Sub-band Graph Encoding-Interaction (SGEI) module, where local graph convolutional networks (GCNs) extract band-specific topological features and lightweight convolutions dynamically allocate contribution weights across bands, followed by a global GCN to capture cross-band interactions; and a time-domain stream that utilizes residual multi-layer perceptron networks to retain local dynamic characteristics of the BOLD signal. To effectively combine these streams, a Feature-Temporal Dual-Dimension Cross-Attention (FTDC) mechanism is introduced, which models both the temporal evolution and inter-domain complementarity for adaptive feature integration. The proposed approach is evaluated on two large-scale neuroimaging datasets—ADNI-2 for mild cognitive impairment (MCI) and ABIDE-II for autism spectrum disorder (ASD)—achieving classification accuracies of 91.40% and 70.18%, respectively.
- 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 dual-stream architecture is well-motivated and elegantly designed to capture both frequency-domain (via wavelet-based sub-band decomposition and GCNs) and temporal-domain (via residual MLPs and LSTM) information, which reflects a deep understanding of the nature of dynamic brain signals.
-
The proposed SGEI module intelligently encodes band-specific features and adaptively integrates them using data-driven weights, effectively addressing a key limitation in previous multi-band fusion methods that relied on fixed or manually defined band weights.
-
The FTDC module introduces a novel two-dimensional attention mechanism that computes attention weights across both the temporal axis and feature streams, providing fine-grained and interpretable feature fusion that boosts model performance.
- The experimental validation is comprehensive, including two challenging classification tasks (MCI and ASD), comparisons with strong baselines (BrainNetCNN, BrainGNN, ST-GCN, Transformer, ACI-FBN), and detailed ablation studies that support the individual importance of each module.
- The paper is clearly structured, and figures like the architecture diagram and performance comparisons are well-executed and 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.
The paper does have a few notable limitations. First, the reproducibility of the method could be improved: there is no mention of code availability or plan to release it, which is especially important given the multi-component architecture and the complexity of the training pipeline. Second, although the model achieves strong performance, the paper could benefit from more in-depth discussion on interpretability, particularly how the attention weights or graph activations correlate with specific brain regions or clinical markers—this would enhance the clinical relevance of the work. Third, while the SGEI module is innovative, its superiority over simpler alternatives is only evaluated through limited ablation (e.g., comparison with attention and concatenation fusion methods), and a deeper comparison with recent graph fusion techniques (e.g., hypergraph or spectral attention-based GNNs) would provide stronger support for its novelty. Fourth, the study relies on nested 10-fold cross-validation but lacks a statistical significance analysis (e.g., p-values) between models to demonstrate that improvements are not due to chance, especially for ASD where margins are smaller. Fifth, Author should mention the full form of SEN, SPE, NC in Table 1. Lastly, computational complexity is not addressed—details on training/inference time or parameter counts would help assess its deployability in clinical settings.
- 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 strong methodological contribution through its dual-stream architecture that effectively integrates frequency and temporal information for dynamic functional connectivity analysis. Its design is well-grounded in neuroscience and machine learning principles, and the introduction of the SGEI and FTDC modules addresses significant limitations in previous multi-band fusion methods. The approach is validated on large, heterogeneous datasets and outperforms strong baselines in both MCI and ASD classification. Although there are some weaknesses in terms of code availability, interpretability, and statistical analysis.
- 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 responses improve confidence in the reproducibility and scientific rigor of the paper, but limitations in novelty analysis and deployment clarity remain. The statistical evaluation is also wrongly addressed in the paper. Most of the work targeted as ‘Future work’ section which is uncovered in the present form. Overall, due to limited novelty, partial experimental justification, and lack of readiness for real-world impact, I will suggest REJECTION of the paper.
Review #2
- Please describe the contribution of the paper
This paper presents a dual-stream multi-band fusion network designed to enhance brain disorder classification from dFC data. The proposed approach aims to better integrate temporal and frequency-specific information through a combination of graph-based modeling and cross-attention mechanisms. While the idea is kind of promising and the method demonstrates encouraging results on two datasets, further evaluation and analysis may be needed to fully assess its generalizability and clinical relevance.
- 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 structure of the paper is very clear, identifying inherent challenges and the limitations of previous approaches. While conventional methods consider multi-band features, they typically rely on simple averaging or fixed-weight strategies to fuse information across frequency bands, failing to adaptively integrate the complementary information among them. This paper addresses these issues by designing a novel architecture. Although constrained by limited space, the experimental design effectively demonstrates the methodological framework.
- 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 methodological equations could be embedded within the paragraphs rather than displayed line by line, which would save considerable space.
The ablation study should include the full set of model parameters used in the experiments to facilitate comparison.
- 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 is generally sound and well-structured, but it would benefit from further strengthening in certain areas.
- 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
1) A dual-stream framework that fuses multi-band connectivity features with spatial topology using attention mechanisms, modeling temporal evolution and cross-domain interactions; 2) A sub-band graph encoding module with local GCNs and global optimization; 3) Superior performance in brain disease diagnosis tasks.
- 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.
Introduces a dual-stream architecture that simultaneously models time-domain dynamics and frequency-domain multi-band connectivity, addressing limitations of prior methods that rely on simplistic averaging or fixed-weight fusion.
- 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 framework does not sufficiently characterize dynamic connectivity patterns across brain regions over time, nor does it identify which specific spatiotemporal features contribute most to disease discrimination. This represents a missed opportunity to reveal clinically meaningful biomarkers and neurobiological insights about disease progression.
- 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 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?
While the proposed method shows promising classification performance, it lacks in-depth spatiotemporal analysis that could significantly enhance its clinical and neuroscientific value.
- 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.
Given the method’s classification performance, computational efficiency, and potential to inspire clinical applications, I recommend acceptance.
Author Feedback
We sincerely thank all reviewers for their constructive comments and acknowledgment of our novel method in addressing the challenges of dynamic functional connectivity analysis for brain disorder classification. Q1: Reproducibility (R1, R2, R3) A1: We will provide a link in the final paper to make code available and ensure reproducibility. Q2: Model Interpretability (R1, R3) A2: We agree with the importance of interpretability analysis and wish to clarify that our framework provides interpretability from multiple perspectives. We can obtain spatio-temporal features from the following two aspects. The FTDC module learns the weights of time windows (Eq. 8) to reveal critical disease-related time segments. And through saliency maps derived from backpropagated gradients, we can analyze the importance of different brain regions. Furthermore, the lightweight 1×1 convolution kernel parameters (Eq. 5) represent the importance of different frequency bands. These spatio-temporal-frequency features highlight which time window, brain region, or frequency band most significantly contributes to disease classification. Due to the limited space in the manuscript, we were unable to elaborate on these aspects. We will include these explanations in the README document of our code repository. Q3: Ablation Study of SGEI Module (R1) A3: As mentioned in the Introduction, many multi-band fusion studies used relatively simple fusion strategies. Researchers either directly applied Concat-Fusion or Attention-Fusion methodologies, or utilized their extended variations. Therefore, in the ablation study of the SGEI module, we compared these two representative baselines. We focused on comparing the entire model with some strong baselines (Table 1), achieving optimal results. The suggestion from R1 is forward-thinking, we will consider comparing with more recent graph fusion techniques in our future work. Q4: Statistical Analysis (R1) A4: We sincerely appreciate the reviewer’s important suggestion, which indeed helps enhance the statistical significance of our findings. On the ACC metric, we have performed paired samples t-test and confirmed that our model shows statistically significant differences (p < 0.05) compared to all baseline methods in Table 1. For example, on the ASD dataset, where performance margins are relatively narrow, our model achieves statistical significance with a p-value of 0.014 when compared to the closest baseline model (ST-GCN). If permitted, statistical significance will be denoted using asterisks (*) appended to performance metrics, with annotations corresponding to predetermined significance thresholds (e.g., *p < 0.05) in Table 1. Q5: Paper Writing (R1, R2) A5: We greatly appreciate the reviewers’ suggestions on some writing issues and tips. To improve clarity, we will include the definitions for the abbreviations as footnotes directly beneath Table 1 in the revised version. Additionally, we will embed short equations (e.g., Eq. 10) into the main text to save space. Q6: Parameter Settings in Ablation Study (R2) A6: The parameter settings in the ablation models are consistent with those of the full model. For example, in the Multi_Freq experiment (Table 2), the time branch is excluded, but the frequency branch remains the same. This clarification will be included in this revision. Q7: Model Computational Complexity (R1) A7: Computational complexity is indeed important in clinical settings. Our model has 1.9M parameters with an average training time of 2.0s per epoch. In comparison, ACI-FBN, which has the closest performance to our model on the MCI classification task, has 995K parameters with an average training time of 2.7s per epoch. And our model outperforms it by 2.52% in ACC (Table 1). Our parameter-augmented architecture preserves computational efficiency and concurrently enhances classification performance. Due to space limitations in the paper, we will allocate a separate section in future work to further discuss it.
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’
The paper is technically sound, with clear descriptions of the methodology, experiments, and results. The use of discrete wavelet transform (DWT) for sub-band decomposition, graph convolutional networks (GCNs) for feature extraction, and lightweight convolutions for dynamic band contribution allocation are well-justified and appropriately implemented. The nested 10-fold cross-validation strategy ensures the reliability of the evaluation, and the comprehensive comparison with existing methods provides a strong benchmark for assessing the performance of DSMFN.
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