List of Papers Browse by Subject Areas Author List
Abstract
Recent evidence suggests that modeling higher-order interactions (HOIs) in functional magnetic resonance imaging (fMRI) data can enhance the diagnostic accuracy of machine learning systems. However, effectively extracting and leveraging HOIs remains a significant challenge. In this paper, we propose MvHo-IB, a novel multi-view learning framework that seamlessly integrates pairwise interactions and HOIs for diagnostic decision-making while automatically compressing task-irrelevant redundant information. Our approach introduces several key innovations: (1) a principled framework combining $\mathcal{O}$-information from information theory with the recently developed matrix-based R'enyi’s $\alpha$-order entropy functional estimator to quantify and extract HOIs, (2) a purpose-built Brain3DCNN encoder designed to effectively utilize these interactions, and (3) a novel multiview learning information bottleneck objective to enhance representation learning. Experiments on three benchmark fMRI datasets demonstrate that MvHo-IB achieves state-of-the-art performance, outperforming existing methods, including modern hypergraph-based techniques, by significant margins. The code of our MvHo-IB is available at \url{https://github.com/zky04/MvHo-IB}.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2025/paper/0646_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to the Code Repository
https://github.com/zky04/MvHo-IB
Link to the Dataset(s)
N/A
BibTex
@InProceedings{ZhaKun_MvHoIB_MICCAI2025,
author = { Zhang, Kunyu and Li, Qiang and Yu, Shujian},
title = { { MvHo-IB: Multi-View Higher-Order Information Bottleneck for Brain Disorder Diagnosis } },
booktitle = {proceedings of Medical Image Computing and Computer Assisted Intervention -- MICCAI 2025},
year = {2025},
publisher = {Springer Nature Switzerland},
volume = {LNCS 15974},
month = {September},
page = {411 -- 421}
}
Reviews
Review #1
- Please describe the contribution of the paper
- This paper introduces Multiview higher-order information bottleneck by integrating nonlinear pairwise FC and higher-order interactions.
- This paper proposed O-information to capture HOIs and introduce the matrix-based Rényi’s α-order entropy estimator for its computation.
- The proposed MvHo-IB outperforms eight other widely used brain network classification methods.
- 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.
- Uses O-information (captures multi-variable interactions beyond pairs) combined with a specific entropy estimator (matrix-based Rényi’s α-order) to quantify HOIs from data.
- A custom “Brain3DCNN” designed to process these HOI features.
- A multi-view IB objective function to learn better representations by combining views and compressing redundancy.
- A multi-view IB objective function to learn better representations by combining views and compressing redundancy.
- 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.
- Calculating the C x C x C O-information tensor is computationally intensive and it is applied on such small datasets. There might be risk of overfitting complex features with small number of subjects.
- While Grad-CAM is used, fully interpreting the C x C x C tensor and pinpointing the exact nature of the crucial three-way interactions remains challenging.
- 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.
(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 paper would be stronger if it explicitly addressed the computational time required for the O-information calculation per subject and discussed the feasibility of scaling the approach to larger datasets. The current presentation leaves a gap between the demonstrated performance on small benchmarks and the potential applicability to the large-scale studies that are increasingly common in the field.
- 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 #2
- Please describe the contribution of the paper
The paper proposes a novel multiview learning framework for brain disorder diagnosis that models high-order interactions (HOIs) in functional connectivity. Specifically, it integrates pairwise and high-order dependencies using O-information theory, and compresses redundant information via a multiview Information Bottleneck (MvIB) strategy. The proposed approach combines a Graph Isomorphism Network (GIN) and a purpose-built Brain3DCNN encoder, enabling the extraction of complementary features from both 2D functional connectivity matrices and 3D brain activity tensors derived from time series. This architecture is designed to yield compact, predictive representations for downstream classification 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.
Major strengths are 1) The integration of O-information with a matrix-based Rényi’s α-order entropy estimator provides a rigorous way to quantify and extract HOIs, which are often overlooked in traditional connectivity modeling.
2) The dual-encoder architecture — using GIN for 2D functional connectivity and Brain3DCNN for 3D temporal features — is well-motivated and tailored to each modality.
- 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 manuscript introduces O-information to model high-order interactions (HOIs) in brain connectivity. However, it lacks a clear justification for preferring O-information over established hypergraph-based methods, which are specifically designed to capture arbitrary n-way relationships among brain regions. Notably, hypergraphs have been effectively utilized in brain network analysis to represent complex associations beyond pairwise interactions . Additionally, the current framework appears to model interactions up to the third order, which may not fully encapsulate the complexity of neural dynamics. The authors should provide a rationale for this limitation and discuss how it impacts the model’s ability to capture the full spectrum of HOIs present in brain connectivity data.
- 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
1) The authors claim to model HOIs, but in practice, the order appears to be restricted to triplets (order 3). The generalization to higher-order settings should be either demonstrated or better qualified.
2) Equation (2) requires revision to clarify the use of parentheses and signs between terms.
- 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 multiview learning framework that integrates O-information theory with a matrix-based Rényi’s α-order entropy estimator to model high-order interactions (HOIs) in functional connectivity. While the approach is innovative and shows promise, the manuscript would benefit from clearer justification for the use of O-information over hypergraph-based methods and a more detailed explanation of the limitations associated with modeling only up to third-order interactions.
- 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 multi-view learning framework called MvHo-IB for diagnosing brain disorders using functional magnetic resonance imaging (fMRI) data. The authors employ the Information Bottleneck principle and O-information theory to capture higher-order interactions, introducing a new Brain3DCNN encoder. The experimental results demonstrate good performance. Overall, the paper presents a certain level of innovation and theoretical contribution, with comprehensive experimental validation. It is worth further consideration.
- 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.
This paper proposes a multi-view learning framework called MvHo-IB for diagnosing brain disorders using functional magnetic resonance imaging (fMRI) data. The authors employ the Information Bottleneck principle and O-information theory to capture higher-order interactions, introducing a new Brain3DCNN encoder. The experimental results demonstrate good performance. Overall, the paper presents a certain level of innovation and theoretical contribution, with comprehensive experimental validation. It is worth further consideration.
- 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.
Overall, the paper has the following issues:
1.Choice of O-information: The authors are encouraged to further discuss the distinction between O-information and other higher-order information metrics, providing more theoretical justification.
2.Parameter Sensitivity Analysis: The paper does not present a sensitivity analysis regarding the Rényi α-order entropy parameter α. It is recommended to add relevant experiments to assess the model’s robustness to parameter changes.
3.Dataset Description: While the paper provides dataset sources, there is insufficient detail on data preprocessing and sample partitioning. More detailed descriptions are suggested.
4.Baseline Selection: Although the authors compare various methods, additional comparisons with the latest Transformer models or other advanced neural networks would be beneficial.
5.Figure Annotations: Some figures, particularly the model architecture diagram, lack sufficient annotations. More detailed descriptions are recommended.
6.Formatting Issues: Some citation formats are inconsistent, such as those on pages [10, 4]. It is suggested to standardize the citation formatting.
- 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?
This paper proposes a multi-view learning framework called MvHo-IB for diagnosing brain disorders using functional magnetic resonance imaging (fMRI) data. The authors employ the Information Bottleneck principle and O-information theory to capture higher-order interactions, introducing a new Brain3DCNN encoder. The experimental results demonstrate good performance. Overall, the paper presents a certain level of innovation and theoretical contribution, with comprehensive experimental validation. It is worth further consideration.
Overall, the paper has the following issues:
1.Choice of O-information: The authors are encouraged to further discuss the distinction between O-information and other higher-order information metrics, providing more theoretical justification.
2.Parameter Sensitivity Analysis: The paper does not present a sensitivity analysis regarding the Rényi α-order entropy parameter α. It is recommended to add relevant experiments to assess the model’s robustness to parameter changes.
3.Dataset Description: While the paper provides dataset sources, there is insufficient detail on data preprocessing and sample partitioning. More detailed descriptions are suggested.
4.Baseline Selection: Although the authors compare various methods, additional comparisons with the latest Transformer models or other advanced neural networks would be beneficial.
5.Figure Annotations: Some figures, particularly the model architecture diagram, lack sufficient annotations. More detailed descriptions are recommended.
6.Formatting Issues: Some citation formats are inconsistent, such as those on pages [10, 4]. It is suggested to standardize the citation formatting.
- Reviewer confidence
Somewhat confident (2)
- [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 provided clear and well-supported responses to the main concerns. Their justification for using O-information is convincing, and the theoretical and biological motivations are appropriate. Minor issues remain but do not affect the overall quality and contribution of the paper. I recommend acceptance.
Author Feedback
Reviewer 1: Our primary goal is to introduce a novel framework for capturing high-order information as an alternative and complement to hypergraph (please see also our reply to R2), rather than to prioritize speed. While computing the full C × C × C tensor is time-consuming, efficient strategies exist to reduce the cost for larger datasets. (1) Under Gaussian assumptions, O-information admits an analytical solution based only on means and variances [12]. (2) There is a fast implementation of the matrix-based entropy functional (MEF) [a]. Although we use the basic MEF in this paper for more precise estimation, both alternatives are viable for larger datasets with minimal accuracy loss.
Interpretability is a challenge for the whole community. Nevertheless, our approach provides new insights that facilitate the identification of informative biomarkers.
Finally, C × C × C tensor is precomputed and stored in advance, so it does not affect the training or inference speed of the network.
[a] Optimal randomized approximations for matrix-based Renyi’s entropy. IEEE Transactions on Information Theory, 2023.
Reviewer 2 [Distinction to Hypergraph] Hypergraph approaches require to manually construct high-order networks by selecting similarity metrics and pruning rules. However, the resulting hyperedges, each connecting an arbitrary number of brain regions, merely indicate that these regions are connected, without revealing how they share information (e.g., redundantly or synergistically).
The O-information sidesteps this ambiguity by providing a single signed measure that indicates if a set of regions generates genuinely new joint information (negative value, synergy-dominated) or primarily reflects repeated signals (positive value, redundancy-dominated).
That is, O-information goes beyond simply indicating if regions are connected. It provides a fine-grained measure that quantifies the nature of their interaction. Moreover, constructing the O-information tensor does not require manually selecting similarity metrics or applying pruning rules that may introduce bias; it is entirely data-driven.
Finally, our O-information outperforms the recent hypergraph approaches [17,13].
[Beyond Third-order] We only discussed third-order O-information in our paper, but the formulation can be easily extended to any higher order (see Eqs.(1)-(3) in [12]), which will result in a K-way tensor.
We chose to use a 3-way tensor for two main reasons. First, introducing a 4-way tensor and extending our framework to three views with an additional 4D encoder (e.g., Brain4DCNN) significantly increased computational cost. Second, the performance gain was marginal.
The terms inside the parentheses in Eq. (2) should be \beta1 I(X1;Z1) + \beta2 I(X2;Z2).
Reviewer 3 [Choice of O-information] We have indeed considered other high-order information-theoretic measures, such as the widely used total correlation (TC) (Eq. (6)) and dual total correlation (DTC) (Eq. (7)), but found their performance to be inferior to that of O-information. From a biological perspective, the use of O-information also aligns with recent trends in neuroscience [23, b].
[b] Emergence of a synergistic scaffold in the brains of human infants. Communications Biology, 2025.
[\alpha = 1.01] We choose alpha=1.01 is just the original authors explicitly recommend α = 1.01 [31,32]. Theoretically, Renyi entropy converges to Shannon entropy as alpha approaches 1, thereby retaining more of the desirable theoretical properties of Shannon entropy, such as additivity and chain rule. Otherwise, O-information does not enjoy elegant expression as in Eq.(5).
[Dataset and baseline] We will add more dataset description. We agree that newer Transformer baselines are informative; however, we deliberately compared against the most recent hypergraph-based networks [17,13], which are more relevant to our paper.
Finally, thanks for pointing out issues of figure and citation; we will revise accordingly.
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