List of Papers Browse by Subject Areas Author List
Abstract
Missing modalities are prevalent in real-world Alzheimer’s disease (AD) assessment and pose a significant challenge to multimodal learning, particularly when the distribution of observed modality subsets differs between training and deployment. Such missingness pattern mismatch induces a conditional representation shift across modality subsets. Existing approaches that rely on implicit imputation or modality synthesis often fail to explicitly model modality availability and uncertainty, leading to overconfident dependence on synthesized features, reduced robustness, and miscalibrated uncertainty estimates. To address these limitations, we propose PRA-PoE, an incomplete multimodal learning framework that is equipped with Prototype-anchored Representation Alignment (PRA) and an Uncertainty-aware Product of Experts (UA-PoE) fusion mechanism. First, PRA uses learnable global proto-types and missing-aware tokens to encode modality availability, distinguish observed from missing modalities, re-synthesize features for missing modalities, and adaptively refine observed representations to align latent spaces across modality subsets, with the goal of reducing representation shift under varying missingness patterns. Second, UA-PoE models each modality as a Gaussian expert and performs closed-form Product of Experts fusion, where experts with higher uncertainty are automatically down-weighted via lower precision, improving uncertainty reliability. We evaluate PRA-PoE under a clinically realistic protocol by training with naturally missing data and testing on all non-empty modality combinations. PRA-PoE consistently outperforms the state-of-the-art across datasets, achieving a 5.4% relative improvement in average accuracy on ADNI and a 10.9% relative gain in average F1 on OASIS-3 over the strongest baseline across all non-empty modality subsets. Code is available at https://github.com/SDH-Lab/PRA-PoE.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/0712_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to the Code Repository
https://github.com/SDH-Lab/PRA-PoE
Link to the Dataset(s)
N/A
BibTex
@InProceedings{YanGua_PRAPoE_MICCAI2026,
author = { Yang, Guangqian AND Du, Ye AND Hou, Wenlong AND Niu, Qian AND Wang, Shujun},
title = { { PRA-PoE: Robust Multimodal Alzheimer’s Disease Classification under Arbitrary Modality Missingness } },
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
Proposed a new approach for learning with missing data, especially missing modalities. Showed a clear improvement to prior 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.
Very nice paper, enjoyed reading it. A new approach, with a well-done evaluation and a good mixture of different experiments and results based on two public datasets, as well as a sufficient description of the given approach.
- 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.
Not much, there are some points which I mostly attribute to the shortness of MICCAI papers, for example 1) the description of tables and figure could be more elaborated, 2) two datasets are used, however, more could be easily added for example using Brats 3) the standard deviation is given, however, other measures of uncertainty like confidence intervals are missing. 4) The discussion is rather short. I wished for more discussing, including limitations of the study as well as limitations of the given approach (e. g. where fellow research might improve it) as well as a general discussion of the topic. 5) Including simple baselines in the evaluation, such as classifiers using only a subset of modalities, mean replacement etc. . . , however, I do not expect that this would change the outcome.
- 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.
(6) Strong Accept — must be accepted due to excellence
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
The paper shows a clear improvement to previous approaches, and a rather complete evaluation (for MICCAI papers).
- 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
This paper addresses multimodal Alzheimer’s disease classification under arbitrary modality missingness, with a particular focus on the mismatch between missingness patterns seen during training and those encountered at deployment. The authors propose PRA-PoE, which combines: (1) Prototype-anchored Representation Alignment (PRA), using learnable modality prototypes, availability-conditioned tokens, and masked cross-attention to align representations across modality subsets and synthesize/refine missing or observed modality features; and (2) an Uncertainty-aware Product of Experts (UA-PoE), which models each modality as a Gaussian expert and performs precision-weighted fusion so that uncertain experts contribute less. The method is evaluated on ADNI and OASIS-3 under a protocol that trains on naturally incomplete data and tests across all non-empty modality combinations. The paper reports consistent gains over several recent incomplete multimodal learning baselines, together with ablations and efficiency measurements, and also provides an anonymized code repository.
- 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 addresses a highly relevant and realistic problem. Missing modalities are common in AD assessment, and the paper correctly emphasizes that modality missingness is not random but protocol- and clinic-driven. Framing the problem as robustness to arbitrary modality subsets is clinically meaningful and important for real-world multimodal AD models.
-
The proposed method is technically coherent and reasonably well motivated. PRA is designed to reduce subset-dependent representation drift through prototype anchoring and availability-aware cross-attention, while UA-PoE introduces an intuitive uncertainty-weighted fusion mechanism by assigning each modality a Gaussian expert and down-weighting high-variance experts through precision aggregation. This is a sensible design for missing-modality settings.
-
The empirical evaluation is fairly strong for a conference paper. The method is compared against several relevant baselines for incomplete multimodal learning. The reported gains are consistent in the summary table and across the listed common modality combinations, and the paper also includes an ablation study and runtime / parameter comparisons.
-
Efficiency is a practical plus. PRA-PoE achieves the strongest reported overall results while remaining relatively lightweight in parameter count and FLOPs, with fast inference speed compared with several baselines.
-
Reproducibility is helped by the anonymized code release.
-
- 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 novelty is good but not fully convincing as a major conceptual advance. The paper combines several existing directions already present in the incomplete multimodal learning literature: modality completion/refinement, prototype or shared latent alignment ideas, and uncertainty-aware / expert-based fusion. The integration is useful, but the contribution feels more like a well-engineered combination of known ingredients than a fundamentally new formulation. The paper itself positions the work against prior lines including shared-specific modeling, imputation-based approaches, M3Care, mmFormer, and recent MoE-based flexible fusion.
2.The uncertainty claim is not validated as directly as it should be. The abstract and method emphasize improved uncertainty reliability and mitigation of overconfident fusion, but the experiments do not report calibration-oriented metrics such as ECE, NLL, Brier score, or uncertainty-quality analyses under different missingness severities. As written, the evidence mainly shows better classification performance, not clearly better uncertainty calibration.
3.The evaluation protocol is clinically motivated, but also somewhat optimistic. The train set contains naturally missing data, yet the test set is built from complete subjects and then modalities are randomly dropped to simulate arbitrary missingness. This is reasonable for controlled benchmarking, but it does not fully demonstrate performance on truly incomplete deployment cohorts, where complete-case subjects may differ systematically from naturally incomplete cases.
4.The paper’s central motivation is robustness to long-tailed and rare modality subsets, but the evidence for this point is incomplete. The authors average over all 15 non-empty subsets and show 8 common combinations, but there is little direct analysis of performance as a function of subset rarity, missingness severity, or mismatch level. That makes the “rare subset robustness” claim less well supported than it could be.
5.Some clarity issues remain. Table 1 is a bit confusing in how datasets/tasks are presented, Eq. (6) contains an unresolved citation placeholder “[?]”, and at least some citations in the introduction appear mismatched to the claim being made. These issues do not invalidate the work, but they reduce confidence in the paper’s polish.
6.Important implementation details are still missing from the main paper. The manuscript states that neuroimaging uses a “3D volume encoder” and tabular data use an MLP, but the exact backbone, latent dimension, and some architectural details are not clearly specified in the main text, which weakens standalone reproducibility even though code is provided.
- 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
The paper tackles an important and realistic problem, and I appreciate the focus on arbitrary missingness rather than assuming fixed modality availability. The empirical gains are promising. The main areas that would strengthen the paper are:
First, please better support the uncertainty-related claims with explicit calibration or uncertainty-quality metrics.
Second, please clarify the split protocol more carefully, especially whether the evaluation may differ between naturally incomplete and artificially masked complete-case test subjects.
Third, because rare modality subsets are central to the motivation, a more direct analysis by subset frequency or missingness severity would strengthen the story.
Finally, please fix the unresolved citation in Eq. (6), clean up citation quality, and clarify key implementation details in the main text.
- 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 lean weak accept on this paper. The topic is important, the proposed framework is well aligned with the practical challenge of arbitrary modality missingness in AD assessment, and the empirical results are consistently strong across two datasets, multiple baselines, ablations, and efficiency comparisons. I also view the use of a realistic train-time missingness protocol plus evaluation across all modality subsets as a meaningful strength.
On the other hand, I do not think the methodological novelty is fully breakthrough-level, and some of the strongest claims, especially around uncertainty reliability and robustness to rare missingness patterns, are not validated as directly as they could be. In addition, the evaluation on complete-case test subjects with synthetic modality dropping, along with some clarity/citation issues, makes me slightly cautious.
Overall, however, I think the paper is above the acceptance threshold and would be of interest to the MIC community, provided the rebuttal addresses the evaluation and presentation concerns.
- Reviewer confidence
Confident but not absolutely certain (3)
- [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.
N/A
- [Post rebuttal] Please justify your final decision from above.
N/A
Review #3
- Please describe the contribution of the paper
The paper introduces a mechanism to handle missing modalities in the classification of Alzheimer disease, using prototype-anchored representation alignment and uncertainty-aware product of experts.
- 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 authors tackle an existing clinical challenge with different modalities being available for different patients.
The authors show clear improvements with their proposed method over baselines and clearly show the contribution of all components in their ablation study.
The methodological contributions are clear and well motivated.
- 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.
There is no real statistical analysis beyond standard deviations. With some values being within the range of the standard deviations, this would help make the significance of the results more clear.
A better choice instead of accuracy would be balanced accuracy, taking into account potential class imbalances.
The paper uses acronyms without defining them properly, e.g. CN, MCI, AD for the target classes in Alzheimer classification.
The description of the datasets should also include class frequencies in both datasets. Especially also regarding the shift between train+val and test set.
- 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 paper is clear and straightforward, however, the proposed solution is very task specific. Together with the above mentioned weaknesses I therefore opt for a weak accept.
- 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
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 realistic problem: learning under missing modalities. A key strength is that training is performed on naturally incomplete data, demonstrating its clinical relevance.
The proposed method is well-motivated and demonstrates consistent improvements over prior work across datasets, modality combinations, and ablations. The experimental evaluation is solid and supports the practical usefulness of the approach.
While the novelty is somewhat incremental and some claims (e.g., uncertainty handling, robustness to rare subsets) are not fully validated, these limitations do not outweigh the paper’s strengths.