List of Papers Browse by Subject Areas Author List
Abstract
Multi-modal disease diagnosis benefits from integrating complementary information across heterogeneous modalities such as medical images, clinical texts, and tabular records. However, in real-world clinical settings, multimodal data are often incomplete, leading to missing modalities that severely degrade the performance of existing methods. Most prior approaches learn decoupled representations in a global feature space, which may overlook class-specific semantics that directly govern the downstream task, resulting in suboptimal cross-modal consistency and discriminative capability. In this paper, we propose Class-aware Semantic Decoupling, a novel paradigm that elevates global feature decoupling to class-level semantic decoupling, enabling explicit decoupling, alignment, and fusion of class-related shared-specific representations in a unified way. Specifically, we present Class-aware Shared-Specific Modeling (CSSM) to derive class-related shared and modality-specific components from multi-modal representations. Building on this, Semantic-guided Decoupling Learning (SDL) performs class-wise cross-modal alignment and shared–specific decoupling, which is guided by class semantic relevance. We further propose Class-aware Shared–Specific Fusion (CSSF) to adaptively integrate shared and specific class semantics, yielding more robust and discriminative predictions. Extensive experiments on image–text and image–tabular multi-label disease diagnosis benchmarks demonstrate that the proposed method significantly outperforms state-of-the-art methods.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/1452_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{ZhoFei_Towards_MICCAI2026,
author = { Zhou, Feixiang AND Xie, Jianyang AND Gao, Zhuangzhi AND Yu, Qinkai AND Li, Jing AND Chen, Long AND Jiang, Zheheng AND Zhao, Yitian AND Meng, Yanda AND Zhao, He AND Lip, Gregory Y. H. AND Zheng, Yalin},
title = { { Towards Class-Aware Semantic Decoupling in Multi-modal Disease Diagnosis with Missing Modalities } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 16885},
month = {September},
page = {pending}
}
Reviews
Review #1
- Please describe the contribution of the paper
This paper proposes Class-aware Semantic Decoupling (CSD) for multimodal disease diagnosis with missing modalties. THe key idea is to move from global shared/specific feature disentanglement to class-level semantic decoupling: class queries are used to extract class-related shared andmodality-specific representations. The shared space is aligned across modalities with semantic-relevance weighting, and a class-aware fusion module combines shared and specific evidence for prediction. The paper evaluates the method on image-text diagnosis with IU-Xray and image-tabular diagnosis with MIMIC under simulated missing-modality settings.
- 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 idea of disentangling disease shared and modality-specific representation is clear. 2.The empirical results are strong, especially under severe missingness and across modality types.
- 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 main contribution of this paper is unclear. The paper writing promites several contributions, including 1) shared and modality-distinct representations, 2) disease-aware attenion fusion, 3) shared-specific features, 4) margin-based redudancy reduction, 5) cross-modal infoNCE and MMD by semantic relevance, 6) query-conditioned masked fusion. But 1) and 2) are introduced in DrFuse, 3) is discussed in ShaSpec, and 5) is similar to Query2Label [1] and CRAL [2]. This makes the novelty of this paper limited on stacking previous previous known ideas. 2.The missingness protocol is synthetic and somewhat unrealistic. The paper defines missing rate by assigning a equal amount of missingness for image-only and aux-only cases. Real clinical missingness is often highly asymmetric. For example, DrFuse notes that in real MIMIC-IV fewer than 20% of patients have chest X-rays. This matters more because the paper’s instance- and distribution-level alignment terms are only defined when both modalities are available, so at high missingness the paired subset driving alignment 3.The empirical reporting does not fully isolate the central claim. Table 2 shows that SDL and CSSF help, but the “baseline” is not clearly defined enough to tell whether the gain comes from class-level decoupling specifically or from a generally stronger attention head. The tables also report single PRAUC values only—no multi-seed variance or confidence intervals—and there is no runtime/memory analysis, even though the method introduces class-wise tensors and losses that scale over disease classes. This again leads to first question, unclear main contribution.
[1] Query2Label: A Simple Transformer Way to Multi-Label Classification [2] Multi-label chest X-ray image classification via category-wise residual attention learning
- Please rate the clarity and organization of this paper
Poor
- 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
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?
With a interesting topic, the paper is unclear about what is the key contribution should be presented.
- 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.
After reading author’s rebuttal, I am happy to raise to accept.
Review #2
- Please describe the contribution of the paper
This paper proposes a Class-Aware Semantic Decoupling framework for multimodal disease diagnosis under missing-modality conditions. By incorporating disease category-level semantic information into modality decoupling, the method effectively learns robust and discriminative multimodal representations. Extensive comparative and ablation experiments further verify the effectiveness of the proposed method.
- 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.This method integrates global modality disentanglement with disease-specific semantics, enabling disease-category-aware feature disentanglement. 2.This method explicitly disentangles features from different modalities and performs both instance-level and distribution-level alignment, with t-SNE visualizations providing evidence of effective disentanglement. 3.The shared and modality-specific features associated with each disease category are adaptively integrated, leading to a more effective fused representation.
- 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 should clarify the necessity of disentanglement. In particular, it would be helpful to explain why disentangled representations are preferable to non-disentangled ones, and whether disease-category-related semantic information could be directly injected into modality features rather than into disentangled features. The motivation for introducing disentanglement should therefore be further strengthened. 2.The main problem addressed in this work is missing modalities in multimodal learning, however, this issue is only briefly discussed in the Introduction and Methodology sections. It remains unclear whether the proposed strategy or framework is specifically designed for missing-modality scenarios. The logical in the Introduction should be further reinforced. 3.In the Semantic-guided Decoupling Learning, the authors define a query for each category, which appears to represent a disease concept. For the classification task, it is worth clarifying whether these disease queries should ultimately remain separated. The authors are also encouraged to provide a t-SNE visualization of the disease queries for better interpretability. 4.In the Methodology section, the notation for each feature should be explicitly marked in Fig. 2, which would make the chapter clearer and easier to follow. 5.Fig. 1(b) is not cited in the main text.
- 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 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?
The proposed method in this paper demonstrates a certain degree of novelty, and the manuscript is generally well organized, with sufficient and reasonably designed experiments. However, the motivation still needs to be further elaborated, and the overall logical flow should be strengthened, particularly to better align with the missing-modality scenario emphasized in the title.
- Reviewer confidence
Very confident (4)
- [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 motivation for decoupling is still not sufficiently clarified and should not be justified merely by the requirements of the setting. Moreover, many multimodal methods without decoupling are still able to effectively learn cross-modal consistency. Therefore, the necessity and advantages of introducing decoupling need to be further emphasized.
Review #3
- Please describe the contribution of the paper
The paper proposes Class-aware Semantic Decoupling (CSD), a multi-modal learning framework for disease diagnosis under missing-modality settings. The key idea is to move from global feature disentanglement to class-level semantic decoupling, where shared and modality-specific representations are explicitly modeled per disease class using a class-query mechanism. The framework introduces: (i) Semantic-guided Decoupling Learning (SDL) to extract class-related shared/specific representations and perform class-aware alignment (instance-level via InfoNCE and distribution-level via MMD), and (ii) Class-aware Shared-Specific Fusion (CSSF) to adaptively fuse these representations for robust multi-label prediction. The method is evaluated on IU-Xray and MIMIC datasets, showing improved PRAUC under varying missing-modality rates.
- 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.
Well-motivated problem setting: Addresses missing-modality scenarios, which are common in clinical practice and highly relevant for real-world deployment. Conceptual shift from global to class-level modeling: The transition to class-aware semantic decoupling is clearly articulated and targets a known limitation of prior disentanglement approaches. Technically structured framework: The method integrates multiple components (class-query attention, redundancy-reduction loss, contrastive alignment, MMD) in a coherent pipeline. Explicit handling of semantic relevance: The weighting of alignment losses based on class-query attention is a meaningful design choice to reduce noise from weakly activated samples. Comprehensive experimental setup: Evaluation includes two heterogeneous benchmarks (image–text and image–tabular), multiple missing-modality ratios, and robustness analysis across train/test mismatch. Ablation study: Systematic removal of components supports the contribution of each module. Consistent performance gains: Improvements over prior methods are reported across all missing ratios, particularly under high missing rates.
- 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.
Limited conceptual novelty over [27]: The method remains within the same shared-specific disentanglement paradigm as [27], with the main difference being the introduction of class-aware structuring. This represents an incremental extension rather than a fundamentally new approach. Limited novelty in individual components: While the overall framework is well-integrated, many elements (cross-attention with class queries, InfoNCE, MMD, shared-specific disentanglement) are established techniques; the novelty mainly lies in their combination and class-level formulation. Lack of statistical significance analysis: Performance improvements are reported without confidence intervals or statistical tests, which is important given relatively moderate gains in some settings. Clinical relevance not deeply analyzed: Although the task is clinically motivated, the paper does not provide insight into clinical impact (e.g., per-class performance, failure cases, or interpretability for clinicians). Handling of missing modalities is simulated: Missing rates are synthetically imposed, which may not reflect real-world missingness patterns (e.g., non-random missingness). No comparison with generative/imputation-based methods under identical settings: Although discussed in the introduction, the experimental comparison does not clearly isolate advantages over imputation-based strategies. Computational complexity not discussed: The addition of class-query attention and multiple alignment losses may introduce non-trivial overhead, but no analysis 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 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
Include statistical significance analysis (e.g., standard deviation across runs or confidence intervals). Include per-class performance breakdown to better assess clinical relevance. Add computational cost analysis (training/inference time, memory) or at least discuss it. Consider comparing with imputation-based baselines under identical experimental settings or at least justify why not.
- 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 addresses a relevant and practical problem: multi-modal diagnosis with missing modalities and proposes a coherent framework that introduces a meaningful shift toward class-level semantic modeling. The method is technically sound and supported by consistent empirical improvements, particularly under high missing rates. However, the level of methodological novelty is moderate, as it largely builds upon existing components, and the experimental validation lacks depth in terms of dataset diversity, statistical rigor, and clinical analysis. The work is solid and potentially useful for the community, but falls short of being a clear top-tier contribution due to limited novelty and incomplete validation.
- 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.
After reading the rebuttal, I maintain my overall positive assessment of the paper. The authors successfully clarified several important points that were previously ambiguous, particularly regarding the distinction between conventional global disentanglement approaches and their proposed class-aware semantic decoupling framework. The rebuttal also clarified the role of the baseline in the ablation study, the rationale behind semantic relevance-guided alignment, and the motivation for disentanglement in missing-modality settings.
In particular, the response adequately explains that the main contribution is not the introduction of entirely new individual components, but rather the reformulation of shared-specific disentanglement, alignment, and fusion within a class-level semantic space. This clarification strengthens the conceptual coherence of the method and better differentiates it from prior global-feature approaches such as DrFuse.
The concerns regarding synthetic missingness remain partially valid, since the experimental setup does not fully reflect realistic asymmetric clinical missingness patterns. However, the authors reasonably justify the use of symmetric missingness as a controlled evaluation protocol consistent with prior work, and the method demonstrates robustness even under very high missing rates.
The paper still presents moderate methodological novelty, as many architectural components are established techniques integrated into a new formulation. Additionally, statistical significance analysis and computational cost evaluation are not included in the current submission. Nevertheless, the work is technically sound, well-motivated, and supported by consistent empirical improvements across datasets and missing-modality settings.
Overall, I believe the rebuttal sufficiently addressed the main concerns to support a weak accept recommendation.
Author Feedback
We sincerely thank all reviewers for their constructive comments. We will revise the paper accordingly and provide more details, as well as release code upon acceptance. [R1:1; R3:1-2] Contribution/class-level decoupling: Our main contribution is a novel class-aware semantic decoupling paradigm for missing-modality diagnosis, where decoupling, alignment, and fusion are moved from the conventional global feature space, e.g., DrFuse, to a class-related semantic space. This allows to better preserve disease-specific semantic structures. Unlike global methods that decouple pooled features and may mix evidence from different diseases, class-level method performs shared/specific decoupling for each disease class, yielding more discriminative and transferable representations under missing modalities. In particular, unlike standard InfoNCE/MMD that align all samples equally, our semantic-guided version performs class-wise alignment weighted by class semantic relevance, reducing noisy alignment and enhancing disease-relevant features. Moreover, unlike conventional query-based methods, CSSF lets each class query fuse shared/specific evidence within its own decoupled semantic space rather than a shared global feature space, leading to more targeted and robust fusion under missing modalities. [R1:2; R3:5] Missingness protocol: Our method is not restricted to symmetric missingness and can be applied to asymmetric settings as long as some paired samples are available for cross-modal alignment. Following existing work, the symmetric protocol is used only for controlled and fair comparison. Notably, Tab. 1 includes severe missing rates up to 90% with only 10% paired samples, where CSD remains robust. [R1:3] Baseline: In Tab. 2, the baseline removes SDL and CSSF, and uses global feature decoupling with masked attention to fuse global shared/specific features, similar to DrFuse. “+SDL” tests class-level decoupling under the same attention module by averaging class-related features, while “+CSSF” further tests CSSF by replacing the attention module. [R1:3; R3:3, 7] Statistical significance analysis and efficiency: We have computed 95% CIs, and our method still shows consistent advantages over competing methods. Compared with DrFuse, the extra cost mainly comes from the class-query-guided attention in SDL, which is lightweight and does not substantially increase inference time. Due to rebuttal constraints, we cannot add new results, but we will include the CIs and discuss computational cost in the final version. [R2:1-2] Disentanglement and missing: CSD is designed for missing-modality diagnosis. Disentanglement is needed because this setting requires both modality-invariant shared evidence for cross-modal transfer when one modality is missing, and modality-specific cues for complementary diagnosis when available. Directly using entangled features may make cross-modal alignment noisy and weaken complementary information. We will strengthen this motivation in the Introduction. [R2:3] Disease queries: Each query extracts disease-related evidence, but strict separation is unnecessary due to disease correlations. We will consider query t-SNE for interpretability. [R2:4-5] Figure clarity: We will revise the figures and related text accordingly. [R3:4] Clinical relevance: Our experiments focus on robustness to missing modalities, a common clinical issue, and show consistent gains on disease diagnosis. We agree that per-class results, failure cases, and clinician-oriented interpretability would better demonstrate clinical impact, and will discuss this limitation in the final version. [R3:6] Imputation: Tab. 1 compares CSD with six SOTA missing-modality methods, showing its effectiveness. We agree generative/imputation comparisons are useful, but this may introduce noise for heterogeneous image-text/tabular data. CSD avoids reconstruction and learns transferable class-level semantics. We will clarify this and consider such comparisons in future work.
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 Reviewers 2 and 3 recommend Weak Accept, praising the consistent performance gains under missing modality settings, Reviewer 1 raises critical concerns regarding limited novelty and unrealistic synthetic missingness protocols. The authors should clearly articulate the specific advantages of class-level decoupling over global methods, address the novelty concerns by differentiating from prior art, and provide statistical significance analysis to support their claims.
- 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.
Althouth the motivation for decoupling is somehow not sufficiently clarified as suggested by one of the reviewers, most reviewers acknowledge the contributions of this submission.
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 paper addresses a relevant and practical problem in multimodal disease diagnosis under missing-modality conditions. Its main contribution is the move from global shared/specific disentanglement to class-aware semantic decoupling, where alignment and fusion are performed in a disease-level semantic space. This is a meaningful formulation, especially for multi-label diagnosis where different diseases may rely on different modality cues. The experiments on IU-Xray and MIMIC show consistent improvements across modality types and missing rates, particularly under severe missingness, and the ablation studies support the contribution of the proposed SDL and CSSF components. The reviewers raised valid concerns about moderate novelty, synthetic missingness, the necessity of disentanglement, and the lack of statistical or computational analysis. The rebuttal clarifies the distinction from global decoupling methods such as DrFuse, explains the role of semantic-relevance-guided alignment, and provides a more convincing rationale for why disentanglement is useful when one modality is absent. The authors also indicate that confidence intervals have been computed and that computational cost will be discussed in the final version. Some limitations remain, especially the controlled missingness protocol and the need for deeper clinical analysis such as per-class results or failure cases. Nevertheless, the method is technically coherent, the empirical gains are consistent, and the rebuttal sufficiently addresses the main ambiguity around the contribution. I therefore support acceptance, while encouraging the authors to include the promised statistical analysis and clearer discussion of limitations in the final manuscript.
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 class-aware semantic decoupling framework for multi-modal disease diagnosis with missing modalities, where shared and modality-specific representations are disentangled, aligned, and fused in a disease-conditioned semantic space to improve robustness under incomplete inputs. The work introduces a meaningful shift from conventional patient-level global disentanglement toward disease-level semantic disentanglement, where alignment and fusion are performed in a class-aware space, and demonstrates generally consistent improvements across heterogeneous benchmarks, missing-modality settings, and multiple recent and relevant comparison methods. While the methodological novelty is moderate and builds upon existing disentanglement and query-based mechanisms, the proposed formulation is technically coherent and reasonably motivated, with supporting experiments providing evidence for the effectiveness of its key components. Although improved cross-modal alignment is not directly measured, this is partially reflected in the representation visualization, coupled with improved downstream performance and ablation studies. Meanwhile, the rebuttal clarified the motivation and distinction from prior global disentanglement methods and addressed several design concerns. Overall, the work appears technically sound and clinically relevant, and slightly above the acceptance threshold.