List of Papers Browse by Subject Areas Author List
Abstract
Resting-state fMRI has become a valuable tool for classifying
brain disorders and constructing brain functional connectivity networks by tracking BOLD signals across brain regions. However, existing models largely neglect the multi-frequency nature of neuronal oscillations, treating BOLD signals as monolithic time series. This overlooks the crucial fact that neurological disorders often manifest as disruptions within specific frequency bands, limiting diagnostic sensitivity and specificity. While some methods have attempted to incorporate frequency information, they often rely on predefined frequency bands, which may not be optimal for capturing individual variability or disease-specific alterations. To address this, we propose a novel framework featuring Adaptive Cascade Decomposition to learn task-relevant frequency sub-bands for each
brain region and Frequency-Coupled Connectivity Learning to capture both intra- and nuanced cross-band interactions in a unified functional network. This unified network informs a novel message-passing mechanism within our Unified-GCN, generating refined node representations for diagnostic prediction. Experimental results on the ADNI and ABIDE datasets demonstrate superior performance over existing methods. The code is available at https://github.com/XXYY20221234/Ada-FCN.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2025/paper/1927_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to the Code Repository
https://github.com/XXYY20221234/Ada-FCN
Link to the Dataset(s)
ADNI dataset: https://adni.loni.usc.edu/
ABIDE dataset: https://fcon_1000.projects.nitrc.org/indi/abide/
BibTex
@InProceedings{XunYue_AdaFCN_MICCAI2025,
author = { Xun, Yue and Xu, Jiaxing and Gao, Wenbo and Yang, Chen and Wang, Shujun},
title = { { Ada-FCN: Adaptive Frequency-Coupled Network for fMRI-Based 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 = {34 -- 44}
}
Reviews
Review #1
- Please describe the contribution of the paper
This paper presents a GCN-based method for brain disorder classification. The adaptive cascade decomposition can learn frequency sub-bands, while frequency-coupled connectivity learning can capture intra- and inter-band interactions within a network. A unified GCN is leveraged to learn node features for classification and diagnosis. This paper also carried out comparative experiments to evaluate the effectiveness of the proposed method.
- 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.
- Interesting work that learns a unified functional network that combines different aspects of frequency bands.
- The performance gains over baselines are significant.
- This paper is well-organized and easy to follow.
- 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.
- More baselines that focus on decomposition should be included, e.g., FEDformer.
- To improve the generalization of the proposed framework, more GCN architectures should be tested to show the effectiveness of the unified functional network. For example, GraphSAGE and GIN[1].
- Since GCNs are easily affected by over-smoothing, it is interesting to include more comparative results on deep GCN architectures, e.g., DeepGCN[2].
[1] Xu K, Hu W, Leskovec J, et al. How Powerful are Graph Neural Networks?[C]//International Conference on Learning Representations. [2]Li G, Muller M, Thabet A, et al. DeepGCNs: Can GCNs go as deep as CNNs?[C]//IEEE/CVF international conference on computer vision.
- 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 authors claimed to release the source code and/or dataset upon acceptance of the submission.
- 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?
See the strengths and 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
- Adaptive Frequency Decomposition: The paper proposes an Adaptive Cascade Decomposer that learns task-relevant frequency sub-bands for each brain region from fMRI data, rather than relying on fixed or handcrafted frequency band definitions. This adaptive approach better captures disease-specific frequency patterns.
- Unified Frequency-Coupled Connectivity Learning: The model integrates both intra-band and cross-band interactions into a single functional network representation. This is achieved through:
- Intra-Band Connectivity via Dynamic Thresholding: Capturing robust connections specific to each frequency band.
- Cross-Band Coupling with Dual-Projection Bilinear Attention: Modeling interactions between different frequency bands.
- Unified Graph Convolutional Network (GCN): A novel message-passing mechanism within a Unified-GCN that processes the unified functional network and generates refined node representations for diagnostic prediction.
- 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.
- Novel Formulation: Adaptive Frequency Decomposition
- Strength: The paper introduces a novel method called the Adaptive Cascade Decomposer that adaptively learns frequency sub-bands from fMRI time series data.
- Why It’s Interesting: Unlike conventional methods that use pre-defined or handcrafted frequency bands, this method learns task-relevant sub-bands specific to each brain region. This is significant because brain functional connectivity can vary significantly across individuals and disease states.
- Impact: This adaptive decomposition enhances the model’s ability to capture disease-specific patterns, thereby improving diagnostic sensitivity and specificity.
- Innovative Cross-Frequency Connectivity Modeling
- Strength: The proposed Frequency-Coupled Connectivity Learning incorporates both intra-band and cross-band connectivity within a unified network.
- Why It’s Interesting: Previous methods often either ignore cross-frequency interactions or use simple fusion methods that do not adequately model complex relationships between different frequency bands. The Dual-Projection Bilinear Attention mechanism used here specifically addresses this by learning asymmetric and band-specific interactions.
- Impact: This holistic modeling approach allows the network to better capture the complex interplay between different frequency bands, which is crucial for understanding neurological disorders.
- Unified GCN for Frequency-Aware Message Passing
- Strength: The paper introduces a Unified Graph Convolutional Network (Unified-GCN) to process the unified frequency-coupled connectivity matrix.
- Why It’s Interesting: Instead of treating frequency bands separately, the model uses a graph convolution framework to integrate both intra- and inter-band connectivity. This novel integration is essential because signals from different frequencies often reflect distinct physiological processes.
- Impact: By leveraging this integrated message-passing mechanism, the model generates more informative node representations, directly leading to better classification performance.
- End-to-End Framework with Adaptive Learning
- Strength: The framework integrates adaptive decomposition, frequency-aware message passing, and cross-frequency alignment seamlessly within an end-to-end model.
- Why It’s Interesting: Most existing methods require manual preprocessing or separately handling frequency-specific and cross-frequency information. The proposed model automates this process efficiently.
- Impact: This integrated, automated approach not only reduces preprocessing efforts but also enhances model generalizability across different brain disorders. In summary, the major strengths of the paper lie in its innovative approach to frequency decomposition, holistic cross-frequency modeling, robust GCN integration, and fully end-to-end architecture. These contributions make the proposed Ada-FCN a valuable model for fMRI-based brain disorder classification.
- Novel Formulation: Adaptive Frequency Decomposition
- 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.
- Inadequate Interpretability of Cross-Frequency Coupling
- Weakness: The Dual-Projection Bilinear Attention mechanism used for cross-frequency coupling lacks sufficient interpretability.
- Details: While the paper claims that cross-band connectivity patterns are modeled, it does not clearly explain how these patterns correlate with known neurophysiological phenomena. Additionally, the resulting connectivity matrices are not thoroughly examined or compared with clinically established biomarkers.
- Recommendation: Providing more qualitative analysis or case studies where the cross-frequency coupling reflects known patterns of neurological disorders would strengthen the model’s interpretability.
- Limited Insight into Potential Biases
- Weakness: The study does not adequately address potential biases related to class imbalance in the ADNI dataset and the limited sample size used from the ABIDE dataset.
-
Details: The ADNI dataset has a class imbalance, particularly with fewer samples for certain diagnostic groups such as SMC and AD. To enhance the credibility of the classification results, presenting confusion matrices alongside accuracy and AUROC would provide a more comprehensive evaluation. Additionally, while the ABIDE 1&2 dataset consists of around 2,000 subjects in total, the study only utilizes a relatively small subset, which limits the generalizability of the findings.
-
Recommendation: Including confusion matrices to demonstrate how well the model performs across different classes in ADNI and expanding the analysis to include a larger portion of the ABIDE dataset would significantly improve the reliability and generalizability of the results.
- Inadequate Interpretability of Cross-Frequency Coupling
- 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 authors claimed to release the source code and/or dataset upon acceptance of the submission.
- 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
- Clarification on fMRI Preprocessing: In fMRI studies, it is common practice to perform band-pass filtering during preprocessing, typically focusing on the 0.009–0.08 Hz frequency range for analysis. It would be beneficial to clarify how you handled preprocessing in this study. Moreover, the statement in the introduction, “Constructing a single, generalized brain functional connectivity network using the entire frequency spectrum can obscure critical” seems somewhat misleading, as the actual issue lies in analyzing only the low-frequency band. Consider revising this statement to better reflect the problem of restricting the analysis to low-frequency signals.
- Figure 1 (MSCD Block Inconsistency): In Figure 1, the MSCD block does not accurately reflect the description in the paper. Specifically, the 1D convolution used for high-frequency components is not depicted, unlike the convolution block for low-frequency components. To maintain consistency with the mathematical formulation, please include the convolution block for high-frequency components in the figure as well.
- Visualization of Optimal K: Since finding the optimal decomposition level K is crucial to your model, it would be helpful to provide a visual representation of how you determined this value. For instance, a plot showing the performance change as K varies would clearly demonstrate why the chosen value is optimal.
- Frequency Differences by K Value: It would be insightful to present how the actual frequency characteristics of the signals change according to the value of K. Furthermore, if the differences in frequency decomposition vary between disease groups (e.g., AD vs. Autism), it could potentially reveal intriguing patterns and enhance the interpretability of the results.
- Inconsistency Between Text and Figure 2 (MCI Group): In the description of Figure 2, you mention that the L1-L2 connectivity of the MCI group is enhanced compared to AD and CN. However, in the figure itself, the L1-L2 connectivity in the MCI group appears almost white, similar to AD and CN, indicating weak connectivity. This inconsistency between the text and the figure should be addressed to maintain clarity and accuracy.
- Color Bar Consistency in Figure 2: To ensure accurate interpretation, please confirm whether all three plots in Figure 2 use the same color bar range. It would be helpful to display the color bar range for each plot together, as discrepancies in color scaling can lead to misinterpretation of the connectivity patterns.
- 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?
- Strengths:
- Novelty and Innovation: The paper introduces a novel framework, Ada-FCN, that adaptively learns frequency sub-bands from fMRI data rather than relying on fixed frequency ranges. This adaptive decomposition effectively addresses the limitations of conventional methods, which often overlook the multi-frequency nature of neuronal oscillations.
- Unified Connectivity Modeling: The Frequency-Coupled Connectivity Learning method effectively integrates both intra-band and cross-band interactions, capturing more nuanced connectivity patterns essential for brain disorder classification.
- Weaknesses:
- Lack of Preprocessing Details: The paper does not clearly specify the preprocessing pipeline, especially regarding frequency filtering, which is a crucial step in fMRI analysis.
- Potential Bias and Limited Dataset Utilization: The evaluation primarily focuses on ADNI and ABIDE datasets. The ADNI dataset has class imbalance issues, and the analysis on ABIDE uses a small subset despite the availability of a larger population. Including confusion matrices and using a larger ABIDE subset would improve the reliability and generalizability of the results.
- Visualization and Inconsistencies: Some visual representations, such as Figure 1 (missing high-frequency convolution) and Figure 2 (color bar range inconsistency), are either incomplete or unclear, which may confuse the readers.
- Exploration of Optimal Parameters: The determination of the optimal K value (number of decomposition levels) lacks clear visualization, making it hard to assess the robustness of the chosen parameter.
- Overall Assessment: The paper presents a significant advancement in the field of fMRI-based brain disorder classification. Its strengths in novel methodological contributions and superior experimental results clearly outweigh its weaknesses. However, addressing the identified limitations—especially those related to data preprocessing, interpretability, and generalizability—would further enhance the paper’s quality. Therefore, the recommendation leans toward acceptance with the expectation that the authors address the aforementioned issues to strengthen the paper before publication.
- Strengths:
- 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 pewsents a novel framework featuring Adaptive Cascade Decomposition to learn task-relevant frequency sub bands for each brain region and Frequency-Coupled Connectivity Learning to capture both intra- and nuanced cross-band interactions in a unified functional network.
- 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 highlights that different neurological disorders may manifest differences in specific frequency bands, which can significantly affect classification performance. This motivation is pretty meaningful and grounded.
- The proposed Adaptive Cascade Decomposer (ACD) allows the model to learn low- and high-frequency components for each ROI. And the Dual-Projection Bilinear Attention (DPBA) module effectively captures cross-frequency interactions during graph construction.
- The method is thoroughly evaluated against 11 baselines and demonstrates good performance on both ADNI and ABIDE datasets, achieving competitive results on ACC and AUROC (e.g., 79.68% ACC on ADNI and 77.89% on ABIDE).
- 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 claims that 1D convolutions can extract both low- and high-frequency features, but how do you justify how this is achieved or how the learned filters differ across frequency bands. And how convolutional parameters are selected or constrained to ensure frequency specificity?
- The visualizations based on the A_unified are difficult to interpret. What’s the meaning of different colors? and what are the weights of the two matrix components? Also, the loss function includes two weighting hyperparameters (λ1 and λ2), but their values and tuning strategy are not reported.
- What’s the total number of trainable parameters or the computational cost of the model? Given the use of multiple frequency-specific convolutions and attention modules, even with K=2, model size may be significant. A complexity analysis would help assess scalability and risk of overfitting.
- While Figure 2 illustrates group-level differences in connectivity, could you highlight specific brain regions contribute most to these differences?
- 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 authors claimed to release the source code and/or dataset upon acceptance of the submission.
- 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 introduces a framework for frequency-aware brain disorder classification. The use of adaptive frequency decomposition and cross-frequency interaction modeling is novel and relevant. The method demonstrates strong performance on two benchmark datasets and outperforms multiple baselines.
Although some implementation details could be clarified—such as parameter settings and model complexity—these do not detract from the overall contribution.
- 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
- 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