Paper Info Reviews Author Feedback Meta-Review Back to top
List of Papers Browse by Subject Areas Author List

Abstract

Alzheimer’s disease (AD) manifests as a continuous progres-sion from normal cognition (NC) through mild cognitive impairment (MCI) to dementia. However, most deep learning approaches reduce this continuum to disjointed classification tasks, largely ignoring dynamic stage transitions. To decode this complex progression, we propose M3AD, a unified framework that jointly addresses three-class diagnosis classifi-cation and diagnosis stage transition prediction using only T1-weighted sMRI. M3AD leverages an interpretable multi-gate mixture of experts architecture, employing specialized routing mechanisms to dynamically capture both diagnosis-specific pathological patterns and shared struc-tural features across the continuum. It further integrates clinical priors (age, sex, eTIV) via adaptive attention fusion to enhance generaliza-tion. M3AD achieves 95.13% accuracy, compared to 90.44% reported by MCLNC under its original experimental setting, and 94.87% for transi-tion prediction. Crucially, analyzing the multi-gate routing reveals dis-tinct expert activation signatures distinguishing stable from progressive MCI, providing a mechanistic basis for individual-level progression risk stratification.

Links to Paper and Supplementary Materials

Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/0438_paper.pdf

SharedIt Link: Not yet available

SpringerLink (DOI): Not yet available

Supplementary Material: Not Submitted

Link to the Code Repository

https://github.com/csyfjiang/M3AD

Link to the Dataset(s)

ADNI dataset: https://adni.loni.usc.edu/data-samples/adni-data/ OASIS dataset: https://sites.wustl.edu/oasisbrains/

BibTex

@InProceedings{JiaYuf_Decoding_MICCAI2026,
        author = { Jiang, Yufeng AND Ding, Hexiao AND Chen, Hongzhao AND Lan, Jing AND Teng, Xinzhi AND Cheng, Gerald W. Y. AND Mao, Yunlin AND Li, Zongxi AND Xie, Haoran AND Yoo, Jung Sun AND Cai, Jing},
        title = { { Decoding the Alzheimer’s Continuum: Interpretable Multi-gate Routing for Diagnosis and Transition Prediction } },
        booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
        year = {2026},
        publisher = {Springer Nature Switzerland},
        volume = {LNCS 16887},
        month = {September},
        page = {pending}
}

Reviews

Review #1

  • Please describe the contribution of the paper

    This paper proposes a unified multi-task learning framework named M3AD to address diagnosis classification and stage transition prediction within the Alzheimer’s disease continuum. The framework employs a Multi-gate Mixture of Experts (MMOE) architecture, which not only captures diagnosis-specific pathological patterns but also shares structural features. Furthermore, the model innovatively integrates clinical priors (age, sex, eTIV) via adaptive attention fusion technology to enhance its generalization capabilities. The framework achieved outstanding performance on a large cohort comprising 12,037 sMRI scans across six datasets.

  • 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. Figure 1 clearly illustrates that the proposed multi-task unified architecture effectively avoids the diagnostic probability conflicts caused by independent OVO classifiers.
    2. The analysis of the dual-gate routing mechanism reveals distinct expert activations for stable vs. progressive MCI, providing a basis for individual risk assessment.
    3. Integrating features like age, sex, and eTIV at Stage 2 has strong clinical relevance and significantly improves diagnostic accuracy.
  • 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 “2D slice + majority voting” strategy lacks justification. The authors must explain why a native 3D network was not used and address the potential loss of 3D spatial context.
    2. It is unclear whether the baselines in Table 1 benefit from the same SimMIM pre-training as the proposed model. The authors should unify these conditions or provide ablation studies to prove the architecture’s inherent superiority.
    3. The methods compared in Table 1 are relatively old. Quantitative comparisons with recent state-of-the-art models (e. g. , from 2024 or 2025) are needed to validate the paper’s novelty.
    4. Despite mentioning a long-tail distribution, the paper lacks specific strategies for handling imbalanced categories. A detailed per-class performance analysis should be added.
    5. The quantitative discussion on cross-cohort validation (Section 3.2) is inadequate. The paper needs to analyze how differing distributions of clinical priors across cohorts impact model generalization.
    6. Figure 2 fails to clearly highlight the core innovations (like the Adaptive Attention Fusion module). Visual cues such as borders or shading are needed to help readers quickly identify structural novelties.
    7. Equations (2) and (3) use code-style terminology (e. g. , Tokenize, Shift_W, DWConv). These must be replaced with standard, formal mathematical symbols.
  • 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 clinical question is sound, the manuscript is complete, and the novelty is good.

  • Reviewer confidence

    Confident but not absolutely certain (3)

  • [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.

    Accept

  • [Post rebuttal] Please justify your final decision from above.

    The authors have responded to my related concerns.



Review #2

  • Please describe the contribution of the paper

    This paper proposes M3AD, a multi-task multi-gate mixture-of-experts framework that jointly performs three-class Alzheimer’s diagnosis (NC/MCI/AD) and diagnosis stage transition prediction (Stability/Progression/Reversion) from T1-weighted structural MRI alone. The key contributions are a unified architecture replacing independent classifiers integration of clinical priors through an adaptive attention fusion, a two-stage training strategy with SimMIM pretraining to initialize expert specialization followed by multitask fine-tuning; and interpretability analysis. The framework is validated on 12,037 scans across six cohorts.

  • 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 makes a compelling argument that diagnosis classification and stage transition prediction are complementary tasks that should be jointly modeled rather than treated independently. The ablation in Table 2 confirms that multi-task learning consistently outperforms single-task variants, lending empirical support to this design choice.

    • Table 2 systematically dissects the contribution of multi-task learning, SimMIM pretraining, Tok-MLP, clinical priors (individually and combined), and fusion stage placement. Each component is shown to contribute meaningfully, and the ablations are well-structured. The finding that eTIV is the most impactful single clinical prior and that Stage 2 is optimal for fusion are informative design insights.

    • The interpretability analysis is compelling. The expert-gate activation shift for AD-aligned experts when comparing stable vs. progressive MCI is a clear, interpretable signal.

    • Good scale and multi-cohort validation. Using 12,037 scans across six cohorts with proper subject-level fold isolation and cross-cohort generalization testing strengthens the claims considerably. The preprocessing is thorough and welldocumented.

    • Code availability.

  • 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 comparisons in Table 1 are not fair. Results for prior methods are cited from their original papers, which used different datasets, splits, and preprocessing.

    • Why 2D slices? AD atrophy is inherently volumetric. Extracting sagittal slices and doing majority voting throws away 3D spatial information. This choice is never justified, and no 3D baseline is compared.

    • Transition prediction metrics are incomplete: despite severe class imbalance (dominance of “Stability”), only aggregate accuracy and F1 are reported. The paper should report per-class metrics (precision, recall, F1) for each of the three diagnosis classes and each transition class. Aggregate metrics hide performance on the clinically most important subgroups.

    • The cross-cohort experiment (Table 1, right) only tests NC vs. AD binary classification, which is the easiest discrimination task. Cross-cohort evaluation for the three-class and transition tasks would be far more informative.

    • Several design choices are unjustified. Why 8 experts with this specific shared/aligned split? What’s the temperature τ? What’s the λ in Eq. 5? Was task loss weighting explored? Masking ratio 0.6, why?

  • 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 paper addresses a clinically relevant problem with a technically reasonable approach. The multi-task formulation is well motivated, the ablation study is thorough, and the interpretability analysis offers genuinely interesting insights into expert routing behavior.

    However, several issues push it below the acceptance threshold.

    The most critical is the unfair comparison in Table 1: mixing results from different experimental protocols while claiming SOTA improvement is misleading and undermines the paper’s central performance claim. The 2D slice-based approach is a significant unexplained limitation for a volumetric neuroimaging task.

  • Reviewer confidence

    Confident but not absolutely certain (3)

  • [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.

    Accept

  • [Post rebuttal] Please justify your final decision from above.

    The authors have consistently responded all reviewers’ questions. Although I still find the 2D approach a limitation of the proposal, I think this work could raise interest to the attendants to MICCAI 2026.



Review #3

  • Please describe the contribution of the paper

    The main contribution of this paper is the proposal of M3AD, a unified multi-task Vision Transformer that leverages a Multi-gate Mixture-of-Experts and adaptive clinical fusion to simultaneously diagnose Alzheimer’s disease and predict its future progression, validated on over 12,000 multi-center MRI scans.

  • 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 paper proposes M³AD, a unified multi-task framework that jointly models the current Diagnosis Stage (DS) and the future DS Transition (e.g., Stability, Progression, Reversion) within a single architecture, rather than treating them as isolated classification problems. 2.The authors evaluate their method on a massive scale for neuroimaging, utilizing 12,037 T1-weighted MRI scans across six independent cohorts. This large-scale, multi-cohort approach, combined with a rigorous 5-fold cross-validation setup that explicitly confines all longitudinal scans of the same subject to a single fold to prevent data leakage, makes the empirical evaluation credible and robust.

  • 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 authors extract 2D sagittal slices from 3D MRI volumes and assign the subject-level diagnosis and progression labels to every single slice during training. However, Brain atrophy in early-stage Alzheimer’s (MCI) is highly localized (e.g., in the hippocampus and entorhinal cortex). Therefore, many lateral or extreme medial sagittal slices contain absolutely no pathological information. Assigning a “Progressive MCI” label to a healthy-looking slice introduces massive label noise during training, which degrades feature learning. 2.The authors claim the superiority of their multi-task framework, but fail to compare it with a vanilla multi-task learning baseline (e.g., a standard ViT with Hard Parameter Sharing using two task-specific linear heads). In Table 1, prior baselines are only evaluated on the diagnosis task, ignoring the transition task. In Table 2, the ‘Multiple Tasks’ setting only validates that training two tasks is better than one within the proposed architecture, but it does not justify the necessity of the complex components (like the multi-gate routing) over a simple multi-task approach.

  • 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.

    (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 overall score is mainly driven by the strong clinical motivation and large‑scale, multi‑cohort evaluation, which demonstrate that the proposed unified framework is empirically robust and potentially impactful for modeling the AD continuum. However, the paper is limited by methodological concerns regarding the absence of a fair comparison against a simple vanilla multi‑task baseline and the noisy 2D slice‑level labeling strategy, so I lean toward a weak accept (borderline) recommendation.

  • 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 reviewers for recognizing M3AD’s unified multi-task design that resolves NC/MCI/AD exclusivity (R1, R3), the component-wise ablations and interpretable expert-routing for stable vs. progressive MCI (R2), the clinical motivation (R3), and the multi-cohort subject-level evaluation (R2, R3). We respond by theme.

[T1] 2D sagittal slices vs. 3D networks (R1.Q1, R2.Q2, R3.Q1) We agree early AD/MCI atrophy is spatially localized and some sagittal slices may carry limited pathological evidence. Our 2D sagittal design is not a claim of 2D superiority, but a resolution-preserving scalability trade-off. As noted in Raptor [1], 3D volumetric modeling incurs cubic-order computational costs and demands datasets orders of magnitude larger than 2D counterparts, motivating 2D-based alternatives. Under comparable memory, 3D training often requires heavy downsampling (e.g., ~80×96×80 in [2]), weakening subtle medial-temporal cues (hippocampal/entorhinal atrophy). Raptor’s Tab. 7 reports that sagittal-only embeddings already achieve competitive performance, while tri-axis (A+C+S) fusion yields further gains. This supports our sagittal-only design as a reasonable single-view choice under our memory and preprocessing constraints, and we acknowledge multi-view fusion as a clear future direction. Regarding label reliability, slice labels are weak supervision derived from subject-level diagnosis/transition labels. All slices and longitudinal scans of one subject stay within 1 fold; predictions use subject-level majority voting, mitigating isolated unreliable slices at inference though not training-time label noise. We will note this limitation in the revision.

[T2] Table 1 and comparison fairness (R1.Q2/Q3, R2.Q1, R3.Q2) Table 1 is a reference comparison (see footnote); the controlled comparison is Table 2, where all variants share identical ADNI splits, preprocessing, and SimMIM pretraining. We will avoid overstating cross-study superiority and distinguish reference from controlled ablation. The Multiple Tasks rows in Table 2 report full M3AD under joint supervision, not a hard-sharing MTL baseline; they isolate the gain of joint over single-task training under the same routing. Hard-sharing MTL removes diagnosis-aligned experts and task gates, so cannot evaluate our routing. [2] has a different objective/protocol; [3] targets histopathology MIL. Neither addresses NC/MCI/AD exclusivity jointly with transition prediction; unified-protocol comparison remains future work.

[T3] Class imbalance and cross-cohort scope (R1.Q4/Q5, R2.Q3/Q4) Class imbalance is handled via weighted cross-entropy (weight 1.0 for majority, higher for rare transitions: NC-to-AD progression, MCI-to-NC reversion). We will add this detail and report per-class precision/recall/F1 for both tasks. For external validation, transition prediction is limited by the lack of standardized longitudinal MCI follow-up labels, and 3-class cross-cohort diagnosis is complicated by heterogeneous MCI definitions. ADNI has more than 50 sites in its cohort study protocol, which provides intrinsic multi-center evaluation for both diagnosis and transition tasks. Extending external validation beyond NC/AD binary diagnosis remains future work.

[T4] Hyperparameters (R2.Q5) τ=1.0/0.5 (pretrain/finetune); λ=0.5 in Eq.(5); 8 experts = 2 shared + 2 NC + 2 MCI + 2 AD; SimMIM mask 0.6; AdamW lr 1e-4/5e-5, wd 0.05, cosine + 10-ep warmup; task-loss 1:1.A full hyperparameter table will be included in the revision.

[Minors] Figure and notation (R1.Q6/Q7) We will revise Fig. 2 to emphasize dual-gate MMoE routing and Stage-2 clinical prior integration, and replace the code-style Tok-MLP/UNeXt notation [4] with formal mathematical notation. We hope our responses address the reviewers’ concerns and thank them for their constructive feedback. Refs: [1] An et al., Raptor, ICML 2025.[2] Ding et al., Denseformer-MoE, IEEE TMI 2025.[3] Li et al., M4, MedIA 2025.[4] Valanarasu & Patel, UNeXt, MICCAI 2022.



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.
    • The paper jointly models current AD diagnosis and future stage transition in a unified multi-task framework. The large-scale evaluation across six cohorts, the subject-level fold isolation, the detailed ablations, and the expert-routing analysis are clear strengths.
    • The comparison in Table 1 is potentially unfair because prior results appear to come from different datasets, splits, and preprocessing pipelines, which weakens the claimed SOTA performance. In addition, the paper does not compare against a simple vanilla multi-task baseline, making it hard to tell whether the complex multi-gate MoE design is necessary.
    • The 2D slice-based strategy needs strong justification. Assigning subject-level diagnosis and transition labels to all sagittal slices may introduce substantial label noise, especially for localized early AD pathology.
    • The comparison is incomplete due to the absence of a 3D baseline and recent 2024/2025 baselines.
    • The authors should further report per-class transition metrics and clarify key hyperparameters and cross-cohort results for the harder multi-class/transition tasks.
  • 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.

    The authors have responded all reviewers’ questions.



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.

    The overall recommendation from the three reviewers, after rebuttal, leans toward acceptance, with R1 and R2 made their final recommendations to Accept and R3 maintaining a borderline Weak Accept. The reviewers acknowledged the strengths of the unified multi-task framework, the large-scale multi-cohort evaluation with rigorous subject-level isolation, and the interpretable expert-routing analysis for distinguishing stable versus progressive MCI.



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.

    The paper presents a well-motivated work for joint AD diagnosis and status transition prediction, evaluated on a relatively large multi-cohort datasets. Implementations in terms of 2D slices hamper the significance of the proposed method.



back to top