List of Papers Browse by Subject Areas Author List
Abstract
Accurate cancer prognosis prediction from multimodal data is critical for personalized treatment planning, yet remains challenging due to modality-specific technical noise and heterogeneous prognostic characteristics. In particular, whole slide images (WSIs) and mRNA expression profiles exhibit an overlooked asymmetry: WSIs provide stable but weak prognostic cues, whereas mRNA data offer strong prognostic relevance at the cost of high technical variability. This discrepancy between technical uncertainty and prognostic importance is not explicitly modeled by existing multimodal survival models, leading to fragile cross-modal fusion. We introduce the Noise-Aware Disentangled Multimodal Survival Network (NADMSurv), a framework that formalizes importance–uncertainty asymmetry as a guiding principle for multimodal representation learning. An importance–uncertainty dual-track module jointly performs feature preselection and patient-level noise quantification to encode heterogeneous prognostic patterns. A noise-aware disentanglement mechanism then separates multimodal representations into shared representations and modality-specific representations. To model high-order patient correlations, hypergraph neural networks are employed, and subjective logic is introduced to generate opinions with uncertainty quantification, ultimately achieving evidence-level fusion through cross-attention and consensus operators. Multi-cohort evaluation on TCGA datasets demonstrates consistent improvements in C-index over state-of-the-art methods. By analyzing modality-specific technical noise characteristics and disentangling the importance-uncertainty asymmetry, NADMSurv establishes a principled paradigm for robust multimodal survival prediction. Code and sufficient information are available at: https://github.com/imicjs/NADMSurv.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/2148_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to the Code Repository
https://github.com/imicjs/NADMSurv
Link to the Dataset(s)
https://portal.gdc.cancer.gov/
BibTex
@InProceedings{MinWen_NoiseAware_MICCAI2026,
author = { Ming, Wenlong AND Ye, Wenbin AND Liu, Mingxin AND Chen, Depin AND Jiao, Yiping AND Xu, Jun AND Wang, Xiangxue},
title = { { Noise-Aware Importance–Uncertainty Disentangled Multimodal Learning for Robust Cancer Survival 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
Authors propose NADMSurv, a noise-aware disentangled multimodal (WSIs+mRNA) survival network. The core design of this network is separating multimodal representations into shared representations and modality-specific representations. Compared to other multimodal models that are limited to simple feature concatenation, NADMSurv’s explicit feature disentanglement modeling can effectively avoid mutual interference between different modalities. Meanwhile, the uncertainty-guided fusion mechanism enhances the model’s ability to resist noise. Authors validate the method on a prediction task and a risk stratification task, demonstrating its state-of-the-art performance.
- 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.
- Methodological soundness: The method designed in this paper is reasonable, with a complete workflow introduction and detailed formula explanations.
- Relatively sufficient experimental validation: Authors conduct extensive experiments on five cancer types using five public datasets, demonstrating the model’s generalization ability. Furthermore, authors conduct a risk stratification experiment, demonstrating the model’s risk stratification.
- 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.
- Logical disconnect between motivation and methodology: In abstract, authors mention that their framework “formalizes importance–uncertainty asymmetry as a guiding principle for multimodal representation learning”. However, when introducing the importance-uncertainty dual-track module, authors only show the different preprocessing steps for WSI and mRNA, but do not explain how these designs are guided by the modality intrinsic importance-uncertainty asymmetry.
- 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 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.
(5) Accept — should be accepted, independent of rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
Overall, the paper is well-written, the method is well-designed, the experiments are relatively comprehensive, and the idea is also informative for other multimodal tasks. Therefore, I consider this a valuable piece of work.
- 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 addressed all of my previous concerns.
Review #2
- Please describe the contribution of the paper
- Modeling of the importance-uncertainty asymmetry in multimodal data.
- Noise-aware disentanglement of representations in multimodal data.
- 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 importance-uncertainty modeling seems novel and experiments support its usefulness.
- Extensive baseline experiments comparing other methods.
- Proper ablations to show the impact of each component in the model.
- Results on several (open) datasets of different cancers.
- 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.
- Lacking motivation for retaining both the lowest-SE and highest-SE genes.
- Unclear formulations and notation, making it hard to follow the equations. Lacking explanation of x , y, h, z_enc, G, B. What is s_WSI in eq 3? Also, what are i and j (i=variable, j=patient)? Also unclear what the noise pseudo labels e are and how they are created. They show up in eq 6 but how are they calculated? All equations should be looked over to make sure explanations are in place.
- Lacking motivation/reference for employing a poisson-binomial hierarchical model for mRNA. Is this the standard approach?
- Fig 1 is a bit unclear and difficult to follow.
- 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 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
- Please increase the font size in Fig 1.+ Fig 1 is hard to follow. It is unclear which title and label a-e belongs where (the extent of the different panels). It should be clarified.
- Lacking reference for FiLM modulation. Please add.
- The loss in eq 15 should not be referred to as log likelihood, as it is the partial log likelihood. Please adjust.
- 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 NADMSurv seems novel and is shown to improve survival prediction. The experiments seem suitable and well executed and support the conclusions of a well working model. The equations/formulations/notations need some work to improve clarity, and some choices need justification.
- 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.
My concerns have been addressed appropriately, and other reviewers’ concerns also have satisfactory responses
Review #3
- Please describe the contribution of the paper
The authors propose to estimate modality-specific noise, provide an architecture for standardization, and create a new representation. They claim this procedure can improve the concordance index (c-index) in survival analysis. The public cohorts are relatively disjoint in terms of matched pairs (mRNA/WSI). Validation of the methodology to improve c-index is a weaker proposition. It is unclear what contributes to the improvement.
- 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.
see comments
- 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.
see comments
- 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 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.
(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?
see comments
- 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
We sincerely thank all reviewers for the valuable comments and for recognizing the clarity of the paper’s structure, the novelty of methodological ideas, and the sufficient experiments. Below, we systematically address the specific comments (R: Reviewer, M: Meta-Reviewer). 1.Motivation and Importance-Uncertainty Asymmetry (R1, R3, M): We thank you for the opportunity to clarify our core motivation. The principle of importance‑uncertainty asymmetry underpins our noise‑aware architecture. mRNA captures genotype, while WSI captures spatial phenotype. Their discrepancies in data topology and clinical applicability create inherent inter‑modal asymmetry in predictive importance and feature uncertainty, which is exacerbated by modality‑specific technical noise. Thus, we designed an active noise‑aware multimodal learning paradigm to explicitly quantify and constrain noise’s adverse impact on each modality’s prognostic capacity. 2.Gene Selection via SE Metrics (R2, M): We retain both lowest- and highest-SE genes to capture the biological duality of consensus and heterogeneity inherent to cancer progression. Low‑SE genes provide stable, consensus‑driven transcriptomic anchors representing consistent patterns across patients, supporting clinical translation (Dharmesh et al., NAR, 2020). In contrast, high-SE genes reflect intra- and inter-tumoral heterogeneity and distinct oncogenic expression patterns. Together, this dual-retention strategy optimizes the delicate equilibrium between robust core features and highly individualized, high-frequency prognostic signals. 3.Technical Noise Modeling and FiLM Calibration (R2, M): We use wavelet decomposition and a Poisson‑binomial model to quantify patient‑level technical noise and generate rigorous pseudo‑labels. The Poisson-binomial hierarchical structure explicitly models heterogeneous capture efficiency across transcripts through its Beta layer, enabling FiLM-based adaptive calibration. The dynamically estimated noise level guides the affine scaling of multi‑modal feature signals. These patient‑derived noise labels serve as weighting factors in our multi‑objective loss, attenuating disruptive noise during representation learning. The Poisson‑binomial model is a robust statistical framework for mRNA technical noise; we will include the relevant formulations and citations in the revised manuscript. 4.Evaluation Metrics (R3): Although the C‑index is a standard survival metric, rank‑based measures alone cannot ensure absolute risk calibration. We therefore incorporated prognostic data to rigorously validate absolute risk assessment, demonstrating clinical efficacy beyond ordinal ranking. 5.Ablation Studies (R3, M): Regarding the ablation framework, we thoroughly analyzed the individual contributions of our architectural components during preliminary experiments. Due to strict page limits, our manuscript only presents module-level results. While we appreciate your insightful suggestions, which will certainly inform our future work, the official guidelines prohibit incorporating new experimental results during the rebuttal phase. 6.Traceability and Citations (R2, M): We will update the bibliography to include foundational citations for the Poisson‑binomial model and FiLM modulation. Our initial citations of recent HGNN applications in bioinformatics inadvertently obscured traceability to the seminal architecture. We will rectify this by adding the original HGNN reference (Feng et al., AAAI, 2019) while retaining the applied ones for research continuity. 7.Clarity and Reproducibility (R1, R2, R3, M): We thank the reviewers for their helpful suggestions. We will improve Figure 1’s readability, explicitly define all mathematical formulations and derivations in the text, and expand the implementation details, including all hyperparameters. Upon acceptance, the full source code will be released to ensure rigorous reproducibility.
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 addresses an important multimodal survival prediction problem and reports encouraging results on five TCGA cohorts. However, several core methodological claims remain insufficiently supported, especially the assumed importance–uncertainty asymmetry between WSI and mRNA. The rationale for retaining both lowest-SE and highest-SE genes, the Poisson-binomial model for mRNA noise, and the noise pseudo-label construction should be clarified in the rebuttal. The current ablation is also too coarse to isolate the contribution of each major component. In addition, the citation basis for the HGNN propagation formula in Eq. 11 appears questionable, since references [9,16] are not the direct sources for the standard HGNN formulation. This has raised doubts about the paper’s rigor.
- 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 is methodologically relevant and reports solid validation on five TCGA cohorts. The rebuttal adequately addressed the main concerns about importance–uncertainty asymmetry, gene selection, mRNA noise modeling, pseudo-label construction, and citation clarity. Remaining issues are mainly presentation, notation, and reproducibility details rather than core validity. The final version should incorporate these clarifications and improve the mathematical exposition.
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.
This paper proposes a noise-aware disentangled multimodal learning framework for cancer survival prediction using WSI and mRNA data. The rebuttal addresses the main concerns regarding the motivation, gene selection strategy, noise modeling, etc. The remaining issues are mainly related to presentation and implementation clarity, and appear addressable in revision.
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 NADMSurv, a noise-aware multimodal survival prediction framework that explicitly models importance–uncertainty asymmetry and disentangles shared and modality-specific representations. Reviewers appreciated the novelty of the proposed formulation, the comprehensive evaluation across multiple cancer cohorts, and the strong experimental results supported by ablation studies. I agree with the reviewers that the method is technically sound and addresses an important problem in multimodal survival analysis. While some concerns were raised regarding the motivation of certain design choices and the clarity of the presentation, the rebuttal satisfactorily addressed these issues. Overall, I find the contribution novel, well validated, and relevant to the MICCAI community, and I recommend acceptance.