List of Papers Browse by Subject Areas Author List
Abstract
Accurate staging of Alzheimer’s disease (AD) from structural MRI is clinically valuable, yet training reliable multi-class models is constrained by limited labels and by the asymmetric risk of diagnostic errors: a two-stage confusion between cognitively normal (CN) and AD is far more harmful than an off-by-one error with mild cognitive impairment (MCI). We present a label-efficient and safety-aware framework that couples (i) a frozen \emph{patient-conditioned} vision–language prior, built by prompting a pretrained CLIP text encoder with demographics and aligning it with CLIP image embeddings from representative MRI slices, (ii) a \emph{scale-preserving} mixture-of-experts (MoE) router that stabilizes expert utilization under small-batch, imbalanced training, and (iii) an EMA teacher for semi-supervised consistency. Across OASIS and ADNI under 10–50\% labeled data, the proposed model consistently achieves the best accuracy and ordinal agreement while reducing mean absolute error and a clinically weighted mis-staging cost. At 10\% labels, we reach QWK 0.806/0.823 on OASIS/ADNI and reduce clinical cost to 0.158/0.143, substantially shifting the clinical-safety frontier. Notably, with only 20\% labels we match the strongest baseline trained with 50\% labels, indicating a (\sim)2.5(\times) reduction in annotation demand.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/5805_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to the Code Repository
https://github.com/dmlucky/Label-efficient-MoE
Link to the Dataset(s)
N/A
BibTex
@InProceedings{DinSha_PatientConditioned_MICCAI2026,
author = { Ding, Shan AND Wu, Fuzhi AND Kong, Youyong AND Shu, Huazhong},
title = { { Patient-Conditioned Vision–Language Priors with Stable Mixture-of-Experts for Label-Efficient Alzheimer’s MRI Staging } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 16880},
month = {September},
page = {pending}
}
Reviews
Review #1
- Please describe the contribution of the paper
This paper proposes a label-efficient framework for ordinal CN/MCI/AD staging from structural MRI that combines three ideas: (1) a patient-conditioned vision–language prior built from a frozen CLIP text encoder prompted by age/sex and a frozen CLIP image encoder applied to representative MRI slices, (2) a scale-preserving, norm-stabilized mixture-of-experts (MoE) head intended to avoid unstable expert usage in low-label, imbalanced training, and (3) an EMA teacher–student semi-supervised learning scheme for leveraging unlabeled 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 is clinically well framed, which I appreciated. Rather than treating AD staging as a standard flat 3-class classification problem, it emphasizes the ordinal nature of CN/MCI/AD and explicitly evaluates clinically important quantities such as MAE, screening sensitivity/specificity, AD NPV, and a weighted clinical cost.
2.The evaluation is consistently strong across two public cohorts and multiple label-scarce settings. In Table 1, the method improves over all listed baselines on both OASIS and ADNI across 10%, 20%, 30%, and 50% label ratios, and the gains are not confined to a single metric. This makes the empirical result more convincing than a single-dataset or single-split success.
3.The paper’s low-label/data-efficiency story is one of its strongest aspects. The reported result that 20% labeled data matches the strongest baseline at 50% labels is practically meaningful in this domain, where labels are expensive and often noisy. The paper also visualizes the clinical-safety frontier in Figure 2, which is a thoughtful way to communicate the joint benefit in ordinal agreement and lower staging deviation.
4.The proposed architecture is conceptually clean and modular. Figure 1 clearly explains how the frozen VLM prior, demographic encoder, MoE routing, and EMA teacher interact.
5.The paper addresses an increasingly relevant direction in AD imaging: using multimodal priors and label-efficient training rather than only larger image-only backbones.
- 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 claimed “vision–language prior” is weaker than the title suggests. The “text” is not a radiology report or even rich clinical text; it is a templated sentence generated from only age and sex. That may still be useful, but it is much closer to a demographic prompt than to a true clinical language signal.
2.Without a strong baseline that uses simple learned age/sex embeddings or direct concatenation of demographics, it is unclear whether the gain comes from CLIP-based cross-modal semantics or simply from adding demographics in a fancy way.
3.The paper is missing the most important ablations. There is no image-only backbone + SSL baseline inside the same framework, no ablation removing the VLM prior, no ablation replacing the proposed stable MoE with a standard classifier or standard MoE, no study of the ordinal regularizer, and no analysis of whether the gains come from the EMA teacher rather than the proposed architectural components.
4.The baseline set is not fully convincing for a SOTA claim. The paper compares against several semi-supervised baselines, but not against some of the most relevant recent AD MRI methods involving multimodal/VLM or MoE designs, such as ADLIP.
5.Reproducibility from the manuscript alone is not yet strong enough. Several core details are missing or underspecified: the exact 3D MRI backbone, CLIP variant, number of sampled slices K, preprocessing steps, weak/strong augmentations, pseudo-label thresholding, EMA momentum, number of experts, hidden dimensions, and the precise form of the ordinal regularizer.
6.The low-label results are reported on single splits with no variance analysis, which is a weakness for a label-scarce paper. At 10–20% labels, performance can vary substantially with the sampled labeled subset. Repeating experiments across several random subset samplings or reporting standard deviations would make the claims much stronger.
7.There is a potentially important protocol ambiguity: the paper states that train/validation/test splits are subject-wise, which is good, but it does not say whether the labeled/unlabeled partition within the training set is also subject-wise. Since ADNI and OASIS can contain multiple scans per subject, having one scan labeled and another unlabeled from the same subject could artificially make the SSL problem easier. This should be clarified explicitly.
8.Because age and sex are injected directly into the prior, the paper should include at least a brief fairness or subgroup robustness analysis. Otherwise it is difficult to know whether the gains reflect improved disease modeling or increased reliance on demographic shortcuts.
- 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 authors claimed to release the source code and/or dataset upon acceptance of the submission.
- 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?
I place this paper slightly above the acceptance threshold. The strongest reasons are the clinically meaningful problem formulation, the consistent low-label gains across two public cohorts, and the fact that the paper evaluates outcomes that matter in practice—ordinal agreement, catastrophic errors, screening behavior, and weighted clinical cost—rather than only top-1 accuracy.
- 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
In this work, the authors propose a deep learning framework for AD prediction (CN, MC, and AD). The framework is label-efficient
- 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.
- Content is well organized and figures are neat and legible
- The work is well-motivated
- The work is sufficiently evaluated against baselines to demonstrate value and superiority
- 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.
- Despite acknowledging domain-specific encoders such as MedCLIP, CLIP is chosen without justification.
- Authors do not indicate if any experiments were performed to compare and evaluate different prompt templates and their impact on performance.
- Vision-language prior: if s captures the alignment between u and t, why is it once again concatenated with u and t and fed to an MLP? Also, what are the implementation details of the MLP such as number of features and activation function used? What is the value of alpha? Were different alpha values swept?
- Concatenating demographic tokens with text tokens which already include demographic information seems redundant and is not well-motivated or justified.
- The work lacks several implementation details such as learning rate, batch size, momentum value, input normalization strategy, and several other such relevant details.
- 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.
- 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
- For slice-to-image CLIP embeddings, the authors state uniformly sampling along the axial axis. Does this include irrelevant peripheral slices?
- 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 work is relevant and valuable. However, in view of the fact that the authors do not make any mention of availability of source code, the lack of major implementation details severely impedes reproducibility. Further more, most design decisions are not explained or justified.
- 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 #3
- Please describe the contribution of the paper
This paper proposes a label-efficient and clinically safe framework for Alzheimer’s disease staging, which mainly consists of (1) a frozen vision-language prior (VLM prior) constructed based on patient demographic information, (2) a mixture-of-experts (MoE) classification head with scale-preserving and norm-stabilized properties, and (3) a semi-supervised learning mechanism incorporating an EMA teacher. On the OASIS and ADNI datasets, this method achieves optimal performance under label ratios of 10%-50%.
- 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 orientation is clear, grasping the core challenges of label scarcity and clinical asymmetry of diagnostic errors, and designing a targeted clinically weighted cost function. 2.Achieving performance comparable to baseline methods with 50% labels using only 20% labels, addressing the issue of high clinical annotation costs. 3.The three components (VLM prior, MoE routing, and EMA semi-supervised learning) form an organic whole and achieve favorable results.
- 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.It is recommended to supplement ablation studies to verify the necessity and synergistic effects of each component. 2.The comparison baselines are primarily methods from before 2023; it is suggested to compare with the latest methods from 2024-2026.
- 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.
- 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?
This paper makes a contribution to MRI-based staging of Alzheimer’s disease. However, the recommended revisions, specifically ablation studies and expanded baseline comparisons, would substantially enhance the manuscript’s completeness and impact.
- 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
Author Feedback
N/A
Meta-Review
Meta-review #1
- Your recommendation
Provisional Accept
- 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 addresses an important and timely problem: label-efficient ordinal AD staging from structural MRI.
- The empirical evidence is overall convincing. The method shows consistent gains across two public cohorts and multiple label ratios, and the result that 20% labeled data can match strong baselines using 50% labels is practically meaningful for medical imaging scenarios where annotations are costly.
- The framework is also conceptually well organized, combining a frozen VLM prior, stabilized MoE head, and EMA teacher-student learning in a coherent way.
- The authors should further clarify key implementation details and strengthen ablations on the VLM prior, MoE head, and EMA teacher.