List of Papers Browse by Subject Areas Author List
Abstract
Alzheimer’s disease (AD), as a progressive neurodegenerative disorder, poses a growing global health threat, making early diagnosis imperative. Multiview brain network (BN) analysis from resting-state functional MRI (rs-fMRI) has emerged as a promising approach, where brain regions and their interactions are modeled as nodes and edges across complementary views. However, existing methods have limitations. First , they rely on single-measure BNs with fixed nodes and edges, potentially insufficient for capturing complex brain interactions. Second , they lack effective separation of view-consistent and view-specific representations, leading to redundancy and reduced generalizability. To address these challenges, we propose a novel Masked Multiview Brain Network Analysis (MMBNA) framework, integrating multi-measure BNs construction, random masking, and disentangled representation learning. Specifically, we first construct multiview BNs via multi-measure connectivity (capturing full/partial/nonlinear correlations) and multi-granularity masking (at node/edge/feature levels), enriching spatio-temporal-topological diversity while preserving semantic similarity. Subsequently, we perform view-consistent representation learning via cross-view masking, and then a disentangling mechanism is introduced to learn a purer view-specific representation to filter out the redundancy from view-consistent representations, resulting in more compact multiview brain representations. Experiments on the ADNI2 subset of the Alzheimer’s Disease Neuroimaging Initiative (ADNI) dataset demonstrate the effectiveness of the proposed method, achieving significant improvements in diagnostic accuracy and interpretability compared to state-of-the-art approaches.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2025/paper/3411_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{MenDeq_MMBNA_MICCAI2025,
author = { Meng, Dequan AND Guo, Jie AND Wang, Junze AND Xi, Xiaoming AND Qiao, Lishan AND Zhang, Limei AND Liu, Mingxia},
title = { { MMBNA: Masked Multiview Brain Network Analysis via Disentangling for Alzheimer’s Early Diagnosis with fMRI } },
booktitle = {proceedings of Medical Image Computing and Computer Assisted Intervention -- MICCAI 2025},
year = {2025},
publisher = {Springer Nature Switzerland},
volume = {LNCS 15971},
month = {September},
page = {480 -- 489}
}
Reviews
Review #1
- Please describe the contribution of the paper
The authors propose a Multiview brain network analysis approach which combines both multiple views of the network and a disentangled representation framework from fMRI scans for the purpose of Alzheimer’s diagnosis. They use the Graph isomorphism network framework and the transformer-based architecture for masked modeling and the representation of multiple views. They also use 3 different measures (Pearson correlation, partial correlation, and mutual information) as brain network measures. This method is demonstrated on the ADNI dataset to classify healthy controls from Alzheimer’s disease.
- 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 ideas of Multiview modeling (and masked modeling) for fMRI have been proposed before. Additionally, other approaches have used similar approaches to this paper to compute network inputs and also to threshold nodes and edges to construct multiple views. However, this paper combines this idea with disentanglement learning. This is the novel component of this paper.
The experimental results show superior classification performance compared to other methods.
- 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 difference between masked-modeling and multi-view modeling is not made explicit. Do the multiple views represent masks? The authors do not define those two terms separately. From reading the paper, it seems both are same.
The motivation for masked-modeling and multi-view modeling is not described. The concept of masking traditionally coming from applications such as super resolution, patch-based analysis and now implemented in terms of autoencoders and large language models do benefit from spatial or temporal locality (although locality is not a prerequisite). This is useful in applications such as computer vision, image processing, and language/text modeling. In other brain imaging studies, masking has been typically performed on whole brain voxels (neighborhood patches) or cortical surface vertices. Although locality is strictly not required for masked-modeling, such models benefit from the dense redundant information in voxels, sequences etc. In this paper, the authors average the time-series signals for a group of voxels belonging to a particular ROI for a set of N ROIs. Thus, the input data is already low-dimensional (116 x 137). Since the node, edge filtering occurs after the graph is constructed, how does masked modeling help in this situation?
The authors mention adaptive masking in the abstract. Was it used?
Similarly, the purpose of disentanglement learning is not clear. Why is it required?
The authors denote the term sparse correlation (SR) to mean partial correlation. However partial correlation by default is not sparse unless additional constraints (scaled Lasso for example) are imposed. The work by Lee, Hyekyoung, et al. “Sparse brain network recovery under compressed sensing.”, which the authors have cited, calculates sparse partial correlations explicitly. Do the authors use this method to compute the sparse correlations? This is not described.
In this paper, the authors mention they implement node-level, edge-level, and feature-level masking. Node-level and edge-level masking is clear. How is feature-level masking implemented? This was not clear.
For a better comparison, this method should be tested against other methods that have used multi-view approaches. For e.g. the paper Ma, Y., Zhang, T., Wu, Z., Mu, X., Liang, X., & Guo, L. (2023, December). Multi-view brain networks construction for alzheimer’s disease diagnosis. In 2023 IEEE International Conference on Bioinformatics and Biomedicine (BIBM) (pp. 889-892) also uses a multi-view approach.
In Experimental results, the authors compute the accuracy for the binary classification case (CN vs AD). However, in the Ma et al. paper the authors perform a granular classification across 3 classes separately. Thus it would be useful for the authors to compare their method against this approach.
- 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 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
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?
There was a confusion about masked modeling vs Multiview construction. The justification for disentanglement learning was not clear. Even if the experimental results outperform other methods, the paper lacks comparison to other similar approaches that have used similar representations, and Multiview constructions. Those approaches are not cited anywhere in the paper. The classification experiment is limited as only one comparison (CN vs AD) was done after grouping subjects with memory complaints, MCI, and AD as one class.
- 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
The main contribution of the paper is the proposal of a novel framework for early diagnosis of Alzheimer’s disease using resting-state functional MRI named MMBNA. This framework addresses the limitations of existing methods, namely lack of effective separation of representations of the brain network, by integrating multi-measure brain networks (BNs) construction, adaptive masking, and disentangled representation learning.
The framework is trained on the ADNI dataset of 563 subjects and uses 5-fold cross-validation (CV) to split the dataset. Evaluation is based on six metrics: Accuracy (ACC), Area under the curve (AUC), Specificity (SPE), Precision (PRE), Sensitivity (SEN), and F1-score (F1). The proposed method is compared with three state-of-the-art (SOTA) single-view FBN analysis models (STGCN, BrainGNN, and Transformer) and two multiview methods (MVS-GCN and Zhang et al.’s method).
An ablation study is conducted to evaluate the contribution of each component by comparing MMBNA with its four variants: (1) Only Consistent Learning, (2) W/O Disentanglement Loss, (3) W/O Multiview Construction, and (4) W/O Multiview and Disentanglement Loss. The proposed MMBNA framework achieves the best performance on the ADNI dataset for ACC, AUC, SPE, PRE, and F1. Experiments on the Alzheimer’s Disease Neuroimaging Initiative (ADNI) dataset demonstrate that the proposed method improves diagnostic accuracy and interpretability compared to state-of-the-art approaches.
- 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 major strengths of the paper are:
Good redaction: The paper is well-written, providing coherent explanations of the proposed methods and results.
Step-by-Step explanation: The proposed solution, the Masked Multiview Brain Network Analysis (MMBNA) framework, is explained step-by-step manner. This explanation helps readers grasp the innovative aspects of the framework.
Training method: The training method is clearly explained, including the data preprocessing pipeline. This level of detail ensures reproducibility.
Literature: The solution is extensively compared with existing state-of-the-art methods in the literature. This comparison demonstrates the strengths and improvements of the proposed framework over existing approaches.
Connectivity Results with Literature: The final connectivity results are compared with those reported in the literature, including the distribution of results.
Weaknesses : The conclusion of the paper acknowledges the weaknesses and areas for improvement of the proposed method, which is appreciated.
- 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 major weaknesses of the paper are:
-
Lack of practical aspect: While the theoretical aspects of the proposed solution are well explained, the paper lacks detailed practical insights into the implementation of the network like specifics on the architecture, hyperparameters, tome to inference.
-
Unclear review of literature: It is not clear how the literature solutions were reviewed and compared. Specifically, the paper does not clarify whether the state-of-the-art models (STGCN, BrainGNN, Transformer, MVS-GCN, and Zhang et al.’s method) were retrained on the same ADNI dataset. This ambiguity makes it difficult to assess the fairness and validity of the comparisons.
-
Disorganized structure: The organization of the article is somewhat disorganized, which can hinder the reader’s ability to follow the flow of information. Specific points of disorganization include:
Example 1 Ablation study: The ablation study should not be in the discussion section but in the results section. Example 2 Performance: The stating on the performance of the MMBNA framework should not be at the end of the materials and methods section but in discussion
-
- 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 has several strong points: it explains the theoretical framework well and clearly outlines the training method, which helps with reproducibility. It also compares the proposed method extensively with existing literature and acknowledges its weaknesses and areas for improvement.
Despite the weaknesses, the paper is clear enough to be understood and the methods detailed enough to be reproduced by other researchers.
- Reviewer confidence
Somewhat confident (2)
- [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
The paper proposes a functional brain connectome representation learning to be utilized in neurodegenerative conditions. The method essentially uses a multi-layered graph representation that incorporates multiple connectivity definitions. These graphs are randomly masked at node, edge and node feature (i.e. Bold signal time span) level to increase variability. This work is an improvement over a prior work where the authors combined masking strategy with view-specific (connectivity specific) representation learning in addition to a common representation learning.
- 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) Sufficiently strong methodological novelty 2) Strong experimental validation including a comparative assessment 3) Significant improvement in diagnostic accuracy for AD, where the learned representations are utilized.
- 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) Poor reproducibility as the details of the method are not complete and code is not shared 2) The subset of the ADNI data used is not sufficiently specified hence although the dataset is open, it is not straight forward to compile the same subset.
- 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 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
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?
This is an active area, the approach pursued is novel and reasonable, esp. when the ill-defined nature of functional connectivity over BOLD signals is considered. Seeing the reported improvement is an encouraging factor for the researchers to move towards multi-dimensional (multi-layered) graph representations.
- 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
Author Feedback
We thank the AC and reviewers for their valuable feedback and their recognition of our “methodological novelty” (R1-R3). Below we address the major concerns. Q1: Reproducibility and missing experimental details (R1&R3) Setup: 5-fold cross-validation; 60 epochs; lr=0.001; batch=32. Hyperparameters: masking ratio=5%; loss weights (α₁=0.6, α₂=β=γ=0.4, λ=0.2). Architecture: 3-layer GIN encoder, 1-layer MLP decoder (dim=128, dropout=0.5), 4-head Transformer encoder. Code: Will release on GitHub upon acceptance. Additional details will be included in the revised manuscript for full reproducibility. Q2: Clarification of ADNI subset (R1) Our analysis uses 563 subjects (CN=154, SMC=165, MCI=145, AD=99 ) from ADNI2, a subset of ADNI. The manuscript will explicitly state the ADNI2 subset usage. Q3: Masked-modeling vs. Multi-view modeling (R2) Masked-modeling: randomly masking (nodes/edges/features of brain network) to improve generalizability via reconstructing the masked components. Multi-view modeling: learning effective brain representation via fusing the masked brain network (BN) multiviews. Q4: Motivation of masking strategy (R2) Our masking strategy enhances multiview diversity and brain representation effectiveness through: 1)Multiview BN Construction: Multi-granular masking (node/edge/feature-level) generates multiple views with spatial-temporal-topological variations while retaining semantics, overcoming traditional multiview limited modeling. 2)Brain Representation Learning: Reconstruction of masked multiviews during view-consistent/specific learning forces the model to recover latent patterns from partial data, enhancing generalization crucial. Q5: Adaptive masking terminology (R2) Our method employs fixed-ratio random masking rather than adaptive masking. The term “adaptive masking” in the abstract was inaccurately and will be revised to “random masking” in the final manuscript. Q6: Purpose of disentanglement learning (R2) Disentanglement learning separates BN representations into view-consistent (shared across views) and view-specific (unique to individual views) components, while eliminating inter-component redundancy. It addresses a key limitation in existing multiview BN analysis, which focuses on common information fusion, neglecting the redundancy information and compromise model generalizability. It will be underscored in the revised manuscript. Q7: Sparse representation (SR) (R2) Default partial correlation lacks inherent sparsity unless constrained by methods like scaled Lasso. In our work, we adopt the approach from [1], computing sparse partial correlations as SR correlation to ensure network sparsity. The manuscript will clarify it further. Q8: Feature-level masking (R2) Feature-level masking randomly perturbs temporal segments of node features (BOLD signal time-series in matrix X). By recomputing correlation matrices from these masked features, this strategy disrupts temporal patterns while retaining core semantics of BN, enhancing robustness. Q9: Extended Comparative Analysis (R2) While our current focus is on binary classification - the popular evaluation paradigm for this dataset [2]- we fully acknowledge the value of multi-class analysis. In the extended manuscript, we will add detailed comparisons with Ma et al.’s multi-view approach and discuss the binary vs multi-class classification. Q10: Practical details and organization (R3) Methodology Fairness: All baseline methods were reimplemented under identical conditions to ensure equitable comparisons. Manuscript Structure: We will restructure sections to enhance logical flow, including relocating the ablation study to align with methodological discussions. [1] Lee H, Lee D S, Kang H, et al. Sparse brain network recovery under compressed sensing[J]. IEEE Transactions on Medical Imaging,2011,30(5):1154-1165. [2] Gan J, Peng Z, Zhu X, et al. Brain functional connectivity analysis based on multi-graph fusion[J]. Medical image analysis,2021,71:102057.
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”.
The paper provides a valuable methodological contribution. I encourage the authors to incorporate reviewer feedback in the final revision, particularly enhancing clarity around masked modeling, better articulating the rationale for disentanglement learning, and providing additional implementation details where feasible. Doing so would strengthen the overall impact and reproducibility of the work.