Abstract

Alzheimer’s disease (AD) disrupts both the spatial organization and temporal dynamics of brain functional connectivity, which can be jointly characterized by fMRI and EEG. fMRI-derived functional connectivity networks (FCNs) encode multiscale spatial topology, and EEG-derived FCNs capture multiband temporal coupling patterns, providing complementary information for AD diagnosis. However, paired fMRI-EEG acquisitions are rarely available in clinical practice, and existing fusion methods require both modalities at inference, limiting real-world deployment. To address this challenge, we propose CMD-AD (Cross-Modality Distillation for AD diagnosis), a three-stage framework that distills fused spatiotemporal knowledge into single-modality encoders to improve their performance. In Stage 1, hierarchical graph attention aggregation (HGAA) encoders are independently pretrained on each modality to extract multiscale fMRI spatial features and multiband EEG temporal features. In Stage 2, scale-wise bidirectional cross-modality fusion (BCMF) integrates fMRI and EEG representations at each parcellation scale, producing fused spatiotemporal targets. In Stage 3, teacher-student distillation transfers the fused spatiotemporal targets back to each modality-specific encoder, enabling accurate single-modality diagnostic performance without paired inputs. Evaluated on 8 multicenter datasets, CMD-AD improves fMRI-only accuracy from 73.51% to 75.50% and EEG-only accuracy from 63.52% to 65.02%, with AUC increasing from 80.88% to 83.70% and from 71.63% to 73.77%, respectively. These results demonstrate that cross-modality training can effectively enhance single-modality deployment for AD diagnosis. Our code is available at: https://github.com/yizhixiaolinya/Code-MICCAI2026_fMRI-EEG.git.

Links to Paper and Supplementary Materials

Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/1174_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{LinXin_CMDAD_MICCAI2026,
        author = { Lin, Xin AND Liu, Yuxiao AND Yu, Wenfei AND Dou, Yulong AND Li, Weiqi AND Tan, Minhui AND He, Yuanzhe AND Cui, Zhiming AND Shen, Dinggang},
        title = { { CMD-AD: Cross-Modality Distillation from fMRI-EEG for Single-Modality Alzheimer’s Disease Diagnosis } },
        booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
        year = {2026},
        publisher = {Springer Nature Switzerland},
        volume = {LNCS 16886},
        month = {September},
        page = {pending}
}


Reviews

Review #1

  • Please describe the contribution of the paper

    This paper focuses on the clinical necessity of multi-modal brain imaging (fMRI-EEG) diagnosis for Alzheimer’s disease (AD), and proposes an innovative three-stage framework named CMD-AD. The methodological framework is well-designed and the experimental data are relatively sufficient to support the innovation.

  • 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 designs a three-stage framework for cross-modal fusion and knowledge distillation to meet the demand for multi-modal diagnosis of Alzheimer’s disease. It achieves efficient complementation between fMRI spatial structures and EEG temporal features, and transfers the capability of the fusion model to the single-modal encoder via distillation. The effectiveness of the proposed method is validated on eight multi-center datasets.

  • 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.Table 2 shows that performance drops significantly when the multi-scale hierarchical structure is removed, yet the correlation between the multi-scale hierarchy and pathological brain region characteristics in AD is not explained.

    2.Experimental comparisons mainly focus on several traditional methods and unimodal models, lacking direct comparisons with representative multimodal state-of-the-art methods in recent years.

    3.It is suggested to add modal weight analysis and interpretable visualization to improve the scientificity and credibility of the proposed method.

  • 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.

  • Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?

    N/A

  • 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/2026/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?

    Main Factors Leading to the Total Score 1.Strong Strengths Supporting Strong Acceptance a.The paper targets AD diagnosis, a global public health challenge with urgent clinical demands. By effectively fusing fMRI (spatial structure) and EEG (temporal features), the work directly addresses the complexity of AD pathology, demonstrating substantial academic value and potential for clinical translation. b.The proposed three-stage cross-modal fusion and knowledge distillation architecture is logically coherent and technically novel. The integration of multi-modal fusion and knowledge distillation ensures both high performance and practical deployability in real-world clinical scenarios, showing clear innovation over single-modal or conventional fusion approaches. c.Extensive experiments on eight multi-center datasets provide strong evidence of the framework’s effectiveness and generalization capability. The results sufficiently validate the core contributions of the work, with adequate ablation studies and experimental controls supporting the reliability of the conclusions. d.The paper is well-organized, with clear explanations of the model architecture, implementation details, and experimental settings. Most key hyperparameters and procedures are well-documented, ensuring good reproducibility of the research. 2.Minor Deficiencies, Not Barring Acceptance a.While the performance drop with the removal of the multi-scale hierarchical structure is observed, the underlying connection to AD pathological brain regions is not fully explained. This is a minor point and can be easily addressed with additional analysis in the revision. b.The comparative experiments focus on traditional and single-modal methods, lacking direct head-to-head comparisons with recent state-of-the-art multi-modal models. However, the existing results already demonstrate clear superiority on the evaluated benchmarks, and this gap can be closed in the rebuttal. c.The paper would benefit from modal weight analysis and visualization of feature fusion to enhance clinical interpretability. However, the core claims of the paper are already well-supported by the experimental data, and this addition would further strengthen the work without changing its overall conclusion.

  • 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 CMD-AD, a cross-modality distillation framework that leverages complementary spatiotemporal information from fMRI and EEG to improve Alzheimer’s disease diagnosis under single-modality settings. Specifically, it introduces a three-stage pipeline—modality-specific hierarchical encoding, scale-wise bidirectional fusion, and teacher–student distillation—that transfers fused multimodal knowledge into individual modality encoders, enabling accurate inference without requiring paired data at test time.

  • 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.
    • Addresses a clinically realistic and important problem: This work enables single-modality deployment while leveraging multimodal information during training, improving clinical feasibility.
    • Comprehensive and convincing evaluation: The method is validated on eight multicenter datasets, including both single-modality and paired cohorts, which is relatively extensive for this domain.
  • 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.The baseline selection is limited and not up-to-date. The comparisons are restricted to conventional single-modality brain network models (e.g., BrainNetCNN, BrainGNN), while more recent and stronger methods such as BrainNetTF [1], Alter [2], as well as multimodal or incomplete-modality learning approaches specifically designed for brain networks, are not included [3, 4, 5]. This raises concerns about the fairness and competitiveness of the reported performance gains. [1] Kan, Xuan, et al. “Brain network transformer.” Advances in Neural Information Processing Systems 35 (2022): 25586-25599.[2] Yu, Shuo, et al. “Long-range brain graph transformer.” Advances in Neural Information Processing Systems 37 (2024): 24472-24495.[3] Wang, Hu, et al. “Learnable cross-modal knowledge distillation for multi-modal learning with missing modality.” International Conference on Medical Image Computing and Computer-Assisted Intervention. Cham: Springer Nature Switzerland, 2023.[4] Yang, Yanwu, et al. “Incomplete learning of multi-modal connectome for brain disorder diagnosis via modal-mixup and deep supervision.” Medical Imaging With Deep Learning. PMLR, 2024.[5] Sun, Kaicong, et al. “Achieving multi-modal brain disease diagnosis performance using only single-modal images through generative AI.” Communications Engineering 3.1 (2024): 96.2.The paper addresses a problem that has been extensively studied in prior work (e.g., incomplete multimodal learning and cross-modal distillation for brain networks). However, the manuscript fails to adequately review these closely related studies, and more importantly, does not clearly articulate what distinguishes the proposed method from existing approaches. As a result, it is difficult to assess the actual novelty and incremental contribution of the work. 3.Several key hyperparameters are introduced but not clearly specified or discussed, including the distillation weights (η, κ), the temperature parameter τ. The paper does not describe how these parameters are selected, nor does it provide their tuning range, step size, or selection criteria, making the method insufficiently transparent and difficult to reproduce.

  • 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.

  • Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?

    N/A

  • 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/2026/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 — marginally below the acceptance threshold, but would not mind if accepted, dependent on rebuttal

  • Please justify your recommendation. What were the major factors that led you to your overall score for this paper?

    The baseline selection is limited and not up-to-date, and several relevant methods are missing. Moreover, although the problem has been widely studied, the paper does not clearly position its contribution against prior work, making the novelty unclear. Overall, I recommend weak reject.

  • Reviewer confidence

    Confident but not absolutely certain (3)

  • [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 rebuttal addresses several missing baselines and implementation details. However, many of these comparisons and discussions were absent from the original manuscript, and the positioning against closely related multimodal/incomplete-modality methods remains insufficiently clear.



Review #3

  • Please describe the contribution of the paper

    The paper proposes CMD-AD, a three-stage cross-modality distillation framework designed for Alzheimer’s disease (AD) diagnosis. It addresses the clinical challenge where paired fMRI-EEG data is rarely available during inference. The method pretrains hierarchical encoders on single modalities (Stage 1) , learns fused representations from paired subjects via scale-wise bidirectional cross-modality fusion (BCMF) (Stage 2) , and distills this fused knowledge back into single-modality encoders (Stage 3). This enables robust single-modality deployment while benefiting from cross-modality training.

  • 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.The method directly targets a highly practical constraint in clinical neuroimaging: training with paired multi-modalities (fMRI-EEG) but deploying with only one. 2.The three-stage design is logically sound. The scale-wise bidirectional cross-attention (BCMF) ensures structured alignment between multiscale fMRI parcellations and EEG band-informed representations. 3.The authors evaluated the framework across eight multicenter datasets. It shows clear quantitative gains against representative single-modality baselines. For instance, after distillation, fMRI-only accuracy improves from 65.25% to 69.75%, and EEG-only accuracy improves from 59.67% to 64.33%. 4.The Grad-CAM saliency maps demonstrate consistent biologically plausible emphasis on association systems across multiple parcellation scales.

  • 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.Stage 2 and 3 training rely on paired fMRI-EEG cohorts, which have a severe lack of AD-positive subjects. Specifically, PEARL-Neuro provides 78 NC pairs with zero AD pairs, while BrainLat provides only 15 AD pairs against 17 NC pairs. This raises concerns about the teacher model’s ability to reliably learn disease-specific spatiotemporal interactions. 2.The submission only compares against standard single-modality FCN classifiers. It lacks comparisons with alternative missing-modality strategies, such as cross-modal synthesis, latent space imputation, or existing KD frameworks tailored for incomplete modalities. Several highly relevant cross-modal knowledge distillation papers are missing from the references. 3.The evaluation involves 8 multicenter datasets with heterogeneous ages and demographics. However, there is no explicit harmonization (e.g., ComBat) or covariate adjustment mentioned. 4.There are minor theoretical/notation inconsistencies. In Eq. (2) and Eq. (4), the text defines channel-aggregated features by averaging over channels, yet the equations utilize both per-channel and aggregated symbols in ways that make the residual injection unclear.

  • 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.

  • Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?

    N/A

  • 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/2026/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 — marginally above the acceptance threshold, but would not mind if rejected, dependent on rebuttal

  • Please justify your recommendation. What were the major factors that led you to your overall score for this paper?

    The recommendation of a Weak Accept balances the method’s strong practical motivation and measured empirical gains against several evaluation gaps. The framework addresses a highly relevant clinical problem—deploying robust models when complementary paired data is unavailable at inference. The ablation studies clearly demonstrate the utility of each component, and overall metrics improve significantly over baselines (e.g., fMRI-only AUC improves from 72.38 to 78.64). However, the paper is held back by the severe lack of AD-positive pairs in the fusion/distillation training split , lack of multi-site harmonization, and the omission of relevant missing-modality baselines. If these evaluation robustness concerns are addressed in the rebuttal, the paper is a solid contribution.

  • Reviewer confidence

    Confident but not absolutely certain (3)

  • [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.

    N/A



Author Feedback

We appreciate all reviewers’ recognition of our technical novelty and practical value. Below, we concisely address the reviewers’ main concerns: 1) R1/R2/R4: clinical motivation and novelty. In clinical practice, paired fMRI-EEG provides complementary spatial-temporal information, but inference often has only fMRI or EEG. CMD-AD targets this bottleneck by leveraging paired fMRI-EEG during training while requiring only one modality at inference. Its main innovation is scale-wise cross-modality distillation: BCMF aligns fMRI and EEG at each scale to construct fused multiscale spatial-temporal features, which are distilled into each single-modality encoder. Compared with HC-GNN/MCSP, CMD-AD discards the fusion branch after training and requires only one modality at inference. It also avoids missing-modality synthesis as in R2-[5]. More importantly, instead of transferring only global latent knowledge (R2-[3,4]), CMD-AD distills per-scale fused features to capture hierarchical brain-network alterations induced by AD. 2) R2/R4: stronger and missing-modality baselines. Under the same subject-level splits and preprocessing, we additionally report R2-[1,2] and R2-[3-5]. R2-[1,2] are competitive at Base (Stage 1), yet CMD-AD exceeds them by 1.7/2.5 and 1.51.0 fMRI ACC/AUC points, and 3.0/5.0 and 2.0/3.0 EEG points. With CMD-AD distillation, R2-[1,2] further gain 4.5/5.2 and 4.7/5.7 fMRI points, and 7.0/8.5 and 6.3/7.5 EEG points. CMD-AD (Distill) also exceeds R2-[3-5] by 3.7/4.3, 2.7/3.7, and 3.7/5.0 fMRI points, and 2.0/3.0, 1.6/2.5, and 3.6/6.0 EEG points. These results show gains also come from scale-wise fusion and distillation, not backbone choice only. 3) R4: paired AD scarcity and Stage-2 sampling. We agree that paired AD subjects are limited, but CMD-AD does not learn the AD/NC boundary from paired data alone. Stage 1 trains disease-aware encoders on all labeled single-modality data. Stage 2 builds on these encoders to learn fMRI-EEG correspondence via BCMF and produce fused representations for Stage 3 distillation. Thus, paired data mainly support cross-modal alignment rather than for classification alone. Class balancing is applied only within the paired training split after subject-level splits are fixed, leaving validation/test distributions unchanged. 4) R2: hyperparameters and reproducibility. We apologize for omit hyperparameter search details. In Stage 3, η and κ weight feature- and logit-level distillation, while τ controls teacher-logit softness. We select them on held-out validation splits using AUC, searching η,κ∈{0.1,0.2,…,1.0} and τ∈{1,2,3,4}. The selected setting, η=0.5, κ=0.5, and τ=2, is fixed for all folds and modalities. We will add these details to Sec. 4.1.5) R1/R4: R1/R4: multiscale pathology and interpretability. We thank the reviewers for requesting clearer pathological interpretation and modal-weight analysis. AD-related FC alterations can occur at both system and regional levels, so coarse parcellations capture system-level disruption while finer parcellations preserve local ROI interactions. This explains why removing the multiscale hierarchy causes the largest performance drop. Fig. 2 supports this interpretation, as Grad-CAM consistently highlights association systems more strongly than primary sensory systems across scales. The learned scale-wise modality weights show greater reliance on fMRI at intermediate ROI scales, reflecting its spatial specificity, while EEG maintains complementary temporal contributions. Additional BCMF feature visualization shows tighter same-class fMRI-EEG alignment and larger AD-NC separation after fusion. 6) R4: harmonization and notation. We apologize for omit harmonization details due to space. FCNs are ComBat-harmonized before training, and all methods use the same harmonized inputs. For Eqs. (2) and (4), H^bar denotes the channel-aggregated feature used to compute attention context, which is then injected into each per-channel feature via the residual connection.




Meta-Review

Meta-review #1

  • Your recommendation

    Invite for Rebuttal

  • Please justify your decision. In case you deviate from the reviewers’ recommendations, explain in detail the reasons why. In case of an invitation for rebuttal, clarify which points are important to address in the rebuttal.

    While the motivation of exploiting multimodal complementarity is meaningful, the current submission has limitations that need to be addressed in the rebuttal including missing related works, evaluations, notations.

  • After you have reviewed the rebuttal and updated reviews, please provide your recommendation based on all reviews and the authors’ rebuttal.

    Reject

  • Please justify your recommendation.

    The paper addresses a clinically relevant problem and proposes a reasonable cross-modality distillation framework for single-modality AD diagnosis. However, the main concerns remain insufficiently resolved after rebuttal, particularly regarding limited novelty relative to existing incomplete-modality and cross-modal distillation methods, inadequate positioning against related work, missing or post-hoc baseline comparisons, and uncertainty about the reliability of the paired fMRI-EEG training data given the severe scarcity of AD-positive paired subjects. Although the rebuttal clarified several implementation details, these issues affect the strength and trustworthiness of the claimed contribution. Rejection is recommended.



Meta-review #2

  • After you have reviewed the rebuttal and updated reviews, please provide your recommendation based on all reviews and the authors’ rebuttal.

    Reject

  • Please justify your recommendation.

    This manuscript is rejected due to multiple critical issues raised by all reviewers. First, the work only compares against traditional single-modality models and standard classifiers, while missing state-of-the-art multimodal, incomplete-modality learning and advanced brain network methods published in recent years, which undermines the fairness and persuasiveness of performance results. Second, the correlation between the core multi-scale hierarchical structure and pathological characteristics of Alzheimer’s disease is not elaborated, and modal weight analysis and reliable visual interpretation are absent. Third, closely related prior studies on incomplete multimodal learning and cross-modal distillation are not thoroughly discussed, and the unique advantages and incremental contributions of the proposed method are not clearly stated. Fourth, key hyperparameters lack detailed specifications, tuning strategies and selection criteria. Additionally, the training datasets suffer from severe imbalance with extremely limited AD-positive samples, raising doubts about the model’s capability to capture disease-related patterns. No effective harmonization or covariate adjustment is conducted for heterogeneous multicenter data, and minor inconsistencies exist in theoretical notations and formula definitions.



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.

    This paper proposes a unified feature distillation from fMRI-EEG for Single-Modality using their multi-site datasets. The reviewers’ critiques were mainly related to performance evaluation with state-of-the-art methods, discussion on experimental setting, and clarification on data harmonization. While those are important aspects, the paper demonstrates its novelty and incorporates intensive data harmization, which maintain its value.



back to top