List of Papers Browse by Subject Areas Author List
Abstract
Mixture-of-Multimodal-Experts (MoME) models have shown strong performance in cancer survival prediction, as different experts can either integrate modalities in a balanced manner or focus on specific features during encoding. However, real-world multimodal cancer data inevitably contain substantial noise, arising from issues such as read errors and missing calls in gene sequencing, as well as uneven preparation and staining of pathological slides. Existing MoME models compute expert weights directly from inputs, which exposes weight computation to input noise, thereby degrading predictive performance and stability. To address this problem, we propose DualGGR-MoME, a Mixture of Multimodal Experts with Dual-Branch Gating and Global-Guided Routing for Cancer Survival Prediction, which incorporates Dual-Branch Gating (DBG) and Global-Guided Routing (GGR) to achieve a noise-robust expert routing. Unlike conventional MoME methods, we decouple the computation of expert weights into two stages: gating and routing, where gating is responsible for generating noise-robust routing signals and routing is responsible for computing the final noise-robust expert weights. Extensive experiments demonstrate that, compared with MoME, DualGGR-MoME improves the C-index by 3.98\%, 5.08\%, and 2.57\% on the TCGA-BLCA, TCGA-LUAD, and TCGA-UCEC datasets, respectively. Noise-robustness analyses further validate the effectiveness of DualGGR-MoME.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/2192_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)
https://portal.gdc.cancer.gov/
BibTex
@InProceedings{TanKai_DualGGRMoME_MICCAI2026,
author = { Tan, Kaiwen AND Feng, Jie AND Zhang, Siqi AND Zhang, Yongbing AND Huang, Yuxin AND Yu, Zhengtao},
title = { { DualGGR-MoME: A Mixture of Multimodal Experts with Dual-Branch Gating and Global-Guided Routing for 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
The main contribution of the paper is the introduction of DualGGR-MoME, a framework that improves multimodal expert models for cancer survival prediction by introducing a dual-branch gating mechanism and a global-guided routing strategy, enabling more robust expert weighting in the presence of noisy inputs.
- 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.
A major strength of the paper is the detailed and well-structured methodological description, which clearly explains the proposed components and their role within the overall framework, contributing to the clarity and solidity of the approach.
Another important strength is the extensive experimental evaluation. The model is assessed across three TCGA datasets and compared against a broad set of baselines, including multiple unimodal models and several multimodal methods. This comprehensive comparison makes the evaluation particularly convincing and provides a solid basis for assessing the effectiveness of the proposed 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.
A weakness of the paper is the motivation regarding noise in genomic data, which is somewhat overstated. The manuscript claims that sequencing errors and missing values are frequent, while in practice these issues can occur but are not necessarily as pervasive as suggested, and this point should be better contextualized.
From a presentation standpoint, several figures are overly dense (e. g. , Figures 1 and 2), making them difficult to interpret. Additionally, the layout could be improved, as multiple tables and figures are concentrated on the final page, which affects readability.
From an evaluation perspective, the lack of statistical significance analysis is a limitation. For example, Table 2 reports performance improvements but does not include p-values or other statistical tests to support the claims, which weakens the strength of the empirical validation.
- 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?
The recommendation of weak accept is mainly motivated by the solid methodological contribution and the clarity of the proposed approach. In particular, the paper provides a detailed and well-structured methodological description, which strengthens the overall quality of the work.
Additionally, the experimental evaluation is thorough, including comparisons with a relatively large number of state-of-the-art models. Although the validation is limited to three TCGA datasets, the breadth of comparative methods considered makes the empirical analysis convincing.
- Reviewer confidence
Very confident (4)
- [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’ response satisfactorily addressed most of the concerns raised during the review process, and the revised version has improved the overall clarity and quality of the work. Since the datasets are already publicly available, I would further encourage the authors to release the source code as well, in order to enhance reproducibility and facilitate future adoption by the community.
Review #2
- Please describe the contribution of the paper
A targeted improvement addressing the issue that “the gating mechanism is prone to noise contamination” in existing multimodal mixture-of-experts survival prediction models. The authors’ starting point is that existing MoME-based methods calculate expert weights directly from the input, whereas pathological images and omics data in the real world contain noise; consequently, the gating mechanism itself is susceptible to contamination, leading to distorted expert weight allocation.
- 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.
First, the problem formulation is reasonable and grounded in practical applications. Pathological images and omics data are indeed susceptible to noise, and if the gating mechanism in multimodal MoE relies directly on noisy inputs, routing distortion may indeed occur. Second, the method design is well-defined; the paper breaks down the contributions into two components—DBG and GGR—and provides a complete routing process through mathematical formulas.
- 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 current ablation study only verifies the number of branches in the dual-branch gating design. It does not isolate several other key design choices, including GGR itself, the concatenation of global context, the signal loss, the entropy loss, and the change from routing based on raw inputs to routing based on expert outputs. 2.Since the method introduces four experts, dual-branch gating, global-guided routing, and bidirectional alternating encoding, the paper should also compare its parameter count, training time, and inference cost against MoME (or relevant MoE-style baseline). 3.The novelty appears to be more of a structural modification to the routing mechanism of existing MoME models rather than a paradigm-level innovation. 4.The central claim of the paper is that expert outputs are less exposed to noise than raw inputs, and therefore produce more robust routing. However, the paper does not provide theoretical analysis, representation visualization, or signal distribution analysis to explain why the outputs should indeed contain less noise.
- 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.
(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?
Overall, the paper presents a meaningful structural refinement of MoME-style routing for multimodal emotion recognition, and the proposed design shows promising empirical performance. However, the current version does not provide sufficient ablation to isolate the contribution of several key components, lacks comparison of parameter and computational costs against relevant MoE/MoME baselines, and offers limited evidence to support its central claim that expert-output-based routing is inherently more robust to noise than input-based routing. In its current form, the work appears more as an incremental architectural modification than a fundamentally novel paradigm, and its overall justification is not yet fully convincing.
- Reviewer confidence
Confident but not absolutely certain (3)
- [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.
1.How can we ensure that the two branches do not learn identical patterns? 2.This work is titled “Mixture of Multimodal Experts,” yet it lacks analysis of the MoE architecture. Typically, experts in MoE have identical structures, under which routing may collapse. Here, the authors design experts with distinct architectures (without clarifying whether these are merely permutations of different structures), which exacerbates the risk of expert failure. How is the degradation of expert functionality — leading to collapse into a single-expert mode — prevented?
Review #3
- Please describe the contribution of the paper
This paper proposes a DualGGR-MoME model for cancer survival prediction, which incorporates Dual-Branch Gating (DBG) and Global-Guided Routing (GGR) to achieve a noise-robust expert routing. Experimental results on three TCGA datasetes demonstrate the superiority and 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.The core contributions, i.e., DBG and GGR, effectively reduces noise exposure by generating robust signals from expert outputs rather than raw inputs. The proposed two-stage design is intuitive, well-structured. 2.Evaluations on three standard TCGA datasets (i.e., BLCA, LUAD, UCEC) using 5-fold cross-validation show consistent and substantial C-index improvements over 11 SOTA methods. 3.The paper is overall well-written and easy to follow.
- 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 proposed method primarily represents an engineering enhancement of the existing MoME framework, achieved by splitting routing stages and incorporating dual-branch and global guidance mechanisms, resulting in limited conceptual novelty. 2.The choice of exactly four fusion experts is not sufficiently justified. In addition, the architectural design of each expert lacks indepth explanation. Comparisons with alternative numbers of experts and different expert structures would help validate the design choices. 3.The paper does not report key efficiency metrics, such as model size, inference latency, training time, or memory consumption. The computational overhead introduced by DBG, GGR, and bidirectional encoding remains unclear, which is particularly important for real-world clinical deployment. 4.For survival prediction tasks, Kaplan–Meier (KM) curves are essential for evaluating the model’s ability to stratify patients into high- and low-risk groups. The absence of KM curve analysis limits the interpretability and clinical relevance of the results.
- 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?
Please see comments above.
- 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.
N/A
Author Feedback
We sincerely thank all reviewers for their time and thoughtful feedback. We deeply appreciate the constructive comments, which help us improve the quality of this work. Below, we clarify several aspects that the reviewers have raised concerns about.
Novelty. We would like to clarify that the contribution of this work goes beyond a combination of existing approaches. We identify a systematic vulnerability in existing MoE-based methods that has not been explicitly addressed: routing weights computed directly from noisy inputs lead to expert weight misallocation and degraded prediction stability. We tackle it by decoupling routing into two stages — gating, which extracts routing signals from expert outputs to avoid direct noise exposure, and routing, which integrates these signals with global representations to compute stable expert weights — with DBG and GGR as their concrete instantiations. We have further elaborated on this distinction in the revised manuscript to make the conceptual contribution more explicit.
Ablation Study Coverage. We would like to clarify that Table 2 does isolate the contribution of key components. The no-branch configuration retains GGR but removes DBG entirely, isolating GGR’s contribution alone (0.680/0.685/0.702 on BLCA/LUAD/UCEC). The single-branch variants isolate each branch’s individual contribution, and the full dual-branch model combines both. This progression jointly validates: (1) GGR alone improves over standard input-based routing; (2) DBG provides substantial additional gains on top of GGR; and (3) dual-branch generation consistently outperforms either branch in isolation. We have added a more explicit clarification of this in the paper to avoid potential misunderstanding.
Computational Cost. DualGGR-MoME introduces modest overhead relative to MoME. On BLCA: parameters increase from 22.26M to 26.50M, training time from 13.6h to 14.1h, and inference time from 0.020s to 0.021s. All additional parameters are confined to the four routing modules, leaving the shared backbone unchanged. Given C-index improvements of 3.98%/5.08%/2.57% on BLCA/LUAD/UCEC and negligible inference overhead, we consider this a favorable accuracy-efficiency trade-off. We have included these efficiency metrics in the revised manuscript for completeness.
Theoretical Justification for Noise Robustness of Expert Outputs. Expert networks inherently attenuate input noise through three mechanisms: (1) RMSNorm normalizes feature magnitudes, reducing sensitivity to scale variations from additive noise; (2) GELU suppresses extreme responses to corrupted inputs; (3) global average pooling aggregates across the full sequence, averaging out localized perturbations. Acting in composition, these operations make expert outputs systematically less sensitive to input noise than raw encoded features. DBG extracts routing signals from these attenuated outputs, and GGR further stabilizes weight computation via global sample representations. Empirically, Fig. 3 provides direct evidence across three datasets, three noise types, and both modalities, consistently showing that DualGGR-MoME maintains stable risk predictions while MoME’s shift sharply as noise intensifies, constituting solid support for the central claim. We will incorporate this theoretical analysis into the final version of the paper.
Number of Fusion Experts and Kaplan-Meier Curves. Both analyses were conducted but could not be included due to MICCAI’s strict page limit. The four-expert configuration was found to be optimal, and KM curve analysis confirms that DualGGR-MoME achieves better patient risk stratification than MoME across all three datasets.
Presentation. We acknowledge that Figs. 1 and 2 are dense and will improve their layout in the final version.
We once again thank all reviewers for their valuable time and feedback.
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.
This paper proposes a multimodal mixture-of-experts framework for cancer survival prediction that uses dual-branch gating and global-guided routing to improve expert weighting under noisy inputs. The reviewers find the problem practical and appreciate the clear method design. However, several concerns remain about the level of conceptual novelty and the justification of the proposed design. The method mainly seems to be a combination of existing paradigms, and the current analysis does not sufficiently isolate the contribution of key components. The reviews also raise concerns about the computational cost analysis, limited support for the claim that expert outputs provide more noise-robust routing signals, etc. The authors are encouraged to address these points in the rebuttal.
- 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 rebuttal addresses several concerns regarding the motivation, component contribution, computational cost, and noise-robust routing. While some concerns remain about the depth of MoE analysis and possible expert behavior, the overall reviewer consensus after rebuttal is positive, and the remaining issues should be addressed in revision.
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 decision is to accept the paper because it presents a highly structured and practical approach to multimodal cancer survival prediction by addressing noise vulnerability in expert routing. During the rebuttal, the authors successfully resolved critical concerns by detailing the minimal computational overhead, clarifying the comprehensive ablation studies, and providing a good theoretical justification for the noise robustness of their architecture. Although one reviewer maintained reservations about potential expert collapse, the extensive empirical validation across multiple datasets and the carefully defended design choices strongly support publication.
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 DualGGR-MoME, a multimodal mixture-of-experts framework for cancer survival prediction with noise-aware routing. Reviewers appreciated the practical motivation, solid methodology, and comprehensive evaluation across multiple datasets and baselines. I agree with Reviewers #1 and #4 that the proposed routing strategy is well motivated and consistently improves performance. While concerns were raised regarding the degree of novelty and the justification of the noise-robust design, the rebuttal satisfactorily clarified these points and provided additional analysis on efficiency and component contributions. Overall, the paper presents a technically sound and well-validated contribution, and I recommend acceptance.