Paper Info Reviews Author Feedback Meta-Review Back to top
List of Papers Browse by Subject Areas Author List

Abstract

Survival prediction is crucial in clinical oncology for cancer prognosis and treatment planning. Recent survival prediction models have shown that integrating pathology whole-slide images (WSIs) and gene expression data effectively improves performance by leveraging complementary morphological and molecular information. However, in clinical practice, complete multimodal data are often unavailable due to high costs, privacy concerns, and workflow constraints, resulting in substantial modality missingness. In such settings, existing methods typically (i) rely on modality-specific or modality-common information shared across patients, often overlooking patient-specific cross-modal dependencies, and (ii) adopt a single static prediction pathway, which limits their ability to adapt to diverse missingness patterns. To address these challenges, we propose Retrieval-based Mixture-of-Experts (RMoE) for robust cancer survival prediction with incomplete multimodal data. To explicitly capture patient-specific cross-modal dependencies, RMoE utilizes bidirectional cross-modal retrieval, identifying phenotypically similar patients and enabling patient-adaptive imputation through biologically relevant neighbors. Furthermore, we introduce a reliability-aware Mixture-of-Experts architecture, where a router dynamically weights specialized experts by assessing retrieval confidence and modality availability, ensuring stable predictions regardless of diverse missing modality settings. Experiments on three public cancer datasets with 60% missingness show that RMoE consistently outperforms current state-of-the-art methods.

Links to Paper and Supplementary Materials

Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/2015_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{LimMin_RetrievalBased_MICCAI2026,
        author = { Lim, Minjoo AND Kang, Bogyeong AND Lee, Hyun Jung AND Jo, Eunjung AND Heo, Keun-Soo AND Liu, Mingxia AND Kam, Tae-Eui},
        title = { { Retrieval-Based Mixture-of-Experts for Patient-Specific Cancer Survival Prediction with Incomplete Multimodal Data } },
        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 paper is well written, clearly structured, and easy to follow. The methodology is thoughtfully designed and introduces a novel approach to handling incomplete multimodal data through the integration of retrieval-based imputation and a reliability-aware mixture-of-experts framework. The overall idea is both interesting and relevant, particularly in realistic clinical settings where missing modalities are common. The proposed method is technically sound, and the motivation is well aligned with practical challenges in multimodal survival prediction.

  • 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 has several strong aspects. First, it proposes a novel formulation for handling incomplete multimodal data through a retrieval-based paradigm, where missing modalities are imputed using patient-specific neighbors rather than global statistics or generative models; this is particularly interesting because it explicitly models patient-level cross-modal dependencies, which are often overlooked in prior work. Second, the integration of a reliability-aware Mixture-of-Experts (MoE) is well-motivated and technically sound, enabling the model to dynamically adapt to different missingness scenarios by weighting unimodal and multimodal experts based on retrieval confidence, which improves robustness in realistic settings. Third, the approach demonstrates strong clinical relevance, as it directly addresses the common issue of missing data in real-world oncology workflows and does not rely on complete multimodal inputs. Fourth, the evaluation is comprehensive, including multiple TCGA cohorts, different modality availability settings, and comparisons with both unimodal and multimodal baselines, showing consistent improvements even under high missingness. Finally, the paper provides supporting analyses such as ablations, retrieval visualization (e. g. , t-SNE), and generalization experiments with alternative backbones, which help validate the effectiveness and flexibility of the proposed framework.

  • Please list the major weaknesses of the paper. Please provide details: for instance, if you state that a formulation, way of using data, demonstration of clinical feasibility, or application is not novel, then you must provide specific references to prior work.

    Despite the strengths of the paper, there are several important limitations.

    First, the degree of interpretability and biological insight is limited. While the proposed retrieval mechanism is motivated by identifying phenotypically similar patients, the paper does not provide sufficient analysis to explain what specific histopathological or genomic features are driving the predictions. The provided t-SNE visualization offers only a high-level view of the embedding space and does not yield clinically actionable insights. In addition, the behavior of the reliability-aware Mixture-of-Experts (e. g. , expert selection patterns and their clinical meaning) is not analyzed, which further limits interpretability.

    Second, the paper does not include clinicopathologic variables such as age, tumor stage, or grade, which are well-established prognostic factors. As a result, it is unclear whether the proposed model captures truly novel multimodal information or whether it is implicitly learning patterns correlated with these clinical variables. This limits the ability to assess the independence and added clinical value of the learned representations. Including such features or providing an analysis/discussion on their potential impact would strengthen the paper’s clinical relevance.

  • 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.

    (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 the weakness section.

  • 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.

    In my view, this manuscript can be accepted.



Review #2

  • Please describe the contribution of the paper

    They propose a bidirectional cross-modal retrieval mechanism that uses the available modality to retrieve phenotypically similar patients from a memory bank in a shared embedding space, and imputes the missing modality via similarity-weighted aggregation of top-k neighbors, with retrieval restricted to the same risk group. They also introduce a reliability-aware MoE architecture, where a router dynamically interpolates between unimodal and multimodal experts based on retrieval confidence, so that the model can fall back to the unimodal expert when retrieval is unreliable. The method is evaluated on three TCGA datasets (BLCA, BRCA, UCEC) under 60% missing.

  • 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 problem is practical. In real clinical workflows, a missing modality is common. The paper targets this relevant practical issue. 2.Compared to prototype-based or shared-representation methods, using patient-specific neighbors to impute missing modalities directly addresses the patient-level heterogeneity argument the authors make in the intro. This idea is well-motivated. 3.RMoE achieves the best performance in most setting, and the observation that genomics-only RMoE on BRCA outperforms the multimodal results of several competing methods is interesting.

  • 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.Lack of baseline. This paper should compare with SOTA methods, such as [1] [1] Zhou, Junjie, et al. “Robust multimodal survival prediction with the latent differentiation conditional variational autoencoder.” CVPR 2025.2.The paper reports results only on BLCA, BRCA, and UCEC. Given that TCGA covers a large number of cancer types and that prior multimodal survival works commonly evaluate on 5 or more cohorts, limiting evaluation to three cohorts makes it difficult to assess whether the method generalizes across cancer types with different tumor biology and survival characteristics. A broader evaluation is needed to substantiate the generality claim. 3.The paper’s core claim is robustness to missing modalities, yet the experiments cover only one missing rate. How does performance change at 30%, 50%, or 80% missingness? Without this, the robustness claim is not fully supported.

  • 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.

    (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?

    The paper addresses a clinically important problem with a reasonably well-motivated design. However, evaluation is limited to only three TCGA cohorts and a single missingness rate. If the authors can extend evaluation to more TCGA cohorts and report results under multiple missingness rates I would be willing to raise the score.

  • 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 authors argue that three cohorts suffice because results are consistent across them. However, the comparison methods they cite (DisPro, ProSurv) themselves evaluate on 5+ TCGA cohorts. Given substantial differences in tumor biology and modality informativeness across cancer types, three cohorts cannot rule out cohort-specific behavior. For missing rates, there is an internal tension in the authors’ defense. They justify the single 60% setting by “following prior methods for fair benchmarking,” yet frame robustness to missing modalities as the paper’s core contribution. These positions conflict: if the protocol just mirrors prior work, the robustness claim cannot exceed what that protocol demonstrates. Substantiating robustness as a headline contribution requires a performance curve across varied missing rates.



Review #3

  • Please describe the contribution of the paper

    The paper’s main contribution is a survival prediction framework for incomplete WSI+genomics data that combines two components:

    Bidirectional cross-modal retrieval to impute the missing modality by retrieving similar patients from a multimodal memory bank using cosine similarity in a learned shared space. Reliability-aware MoE routing that adaptively mixes unimodal and multimodal experts according to a confidence score derived from retrieval similarity.

    A secondary contribution is the use of a risk-aware cross-modal contrastive loss, where patients in the same discretized survival-risk group are treated as positives when learning the retrieval space.

  • 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 problem is important and clinically relevant. The paper directly addresses incomplete multimodal data, which is a real issue in pathology and genomics workflows, rather than assuming full modality availability. This is also a trending topic in multimodal medical learning at the moment.

    The method is conceptually clean and does not rely on generative imputation. The idea is kind of similar to finding a “digital twin” given that we can not reliably get image counterfactuals.

    Analysis is intuitive and supports the authors’ story that both retrieval and reliability-aware expert routing matter.

    The empirical results are quite strong and I do believe these are generalizable gains (as opposed to generative imputation methods).

  • 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.Great work, but really need to stress the clinical contribution better (maybe you can’t change a whole lot this submission, but if you get accepted/later present this work elsewhere, u need to think again about the story logic). My understanding of your greatest contribution is that you get rid of the prevalent generative imputation methods, which real clinicians will without doubt discard in real diagnostics workflows. Your method is not only novel in the technical sense, it is well-motivated by real-life clinical practice (the doctors look at similar patients, and would ideally desire a “ digital twin “ when they think of AI for missing modalities. As someone who work with patients, doctors would buy your method a lot more than a “diffusion for missing data imputation” paper.

    1. MICCAI doesnt encourage more experiments. But I believe a small experiment that is not MCAR could greatly strengthen the story. Missingness often depends on site, workflow, indication, or patient subgroup. If you read the other recent MoE and missingness type of work, which I highly encourage you to compare with (not actually running them , just talkin about it ) and discuss in the related works, you would find out that they are treating different missing patterns as different groups. By doing so, my understanding is that they implicitly encode MNAR subgroup information, although they are not perfect – even within the same subgroup, there could be different missing reasons and the missing modalities could be of high or little clinical value. Therefore it would be really great for you to run something that isnt MCAR and is well controlled and see how it goes.

    Two papers that leverages the “missingness subgroup” mindset that I mentioned above: nips 24 Flex-MoE: Modeling Arbitrary Modality Combination via the Flexible Mixture-of-Experts cvpr 26 REMIND: Rethinking Medical High-Modality Learning under Missingness–A Long-Tailed Distribution Perspective

    3.A potential concern about outcome-informed retrieval space design is that the retrieval learning explicitly uses discretized survival-risk labels to define positives, and training-time retrieval is restricted to same-risk-group entries. Can you talk more about whether this creates optimism?

  • 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.

    (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 multimodal fusion with missingness subfield literally needs papers like this that gets rid of fancy imputations. There is a irreducible information gap for missing modalities and imputation of modalities will literally just inject more noises. This work does the opposite and is well motivated by clinical reality that doctors look for “patient twins” given we can never get the counterfactual data of tests not done.

    Authors also manage to combine this intricately with recent moe architectures elegantly.

    I gave a 4 instead of 5 or 6, because this work can be truly impactful if the not MCAR setting can be discussed more (this is also what other recent missingness works have shifted towards explicitly or implicitly).

  • 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.

    authors have addressed my comments fully, thus I recommend for accept



Author Feedback

We greatly appreciate the constructive feedback from the AC & Reviewers. We respond to all concerns below. R1: ADDITIONAL INTERPRETABILITY ANALYSIS. – We note that analyses of our main component were prioritized. Since our primary goal is to compensate for missing modalities through retrieval, we focus on verifying retrieval effectiveness through t-SNE. Our MoE is interpretable by design, as gating weights reflect confidence in each expert according to retrieval quality, providing transparent decision-making. Additional analyses (e.g., WSI heatmaps and genomic importance scores) were omitted due to limited space. R1: INTEGRATION OF CLINICAL VARIABLES. – Since our framework is modality-agnostic, clinical variables can be naturally incorporated as auxiliary metadata. We agree that this integration may further improve predictive performance. We will discuss it as a potential extension in the final version. R2: ADDITIONAL BASELINES & DATASETS. – We included 8 competitive baselines spanning unimodal, multimodal, and 5 recent methods handling missing modalities during both training and test. – We appreciate suggesting LD-CVAE. While LD-CVAE assumes complete multimodal data at training and specifically addresses missing genomics at test, our RMoE considers a more challenging setting where either WSI or genomics may be missing during both training and test. Therefore, we prioritized methods more closely aligned with our challenging setting. Notably, DisPro (CVPR2025) and ProSurv (MICCAI2025) target survival prediction under all possible missing settings. – We evaluated our RMoE on widely used BLCA, BRCA, and UCEC datasets, which represent diverse cancer types with different cross-modal characteristics. Consistent improvements across all cohorts support the generalizability of our approach. As noted in the Conclusion, we are extending our method to larger external clinical cohorts. R2: MISSING RATES. – Our robustness claim refers to maintaining strong performance under severe missing modality settings during both training and test, unlike many previous methods that only consider test-time missingness. – In this work, we adopt the 60% missingness setting in training, following comparative methods for fair benchmarking. DisPro explicitly evaluates under 60% missingness, while ProSurv reports 60% as the most severe setting in its ablation study. We adopt this severe missing rate during training, expecting that superior performance under such severe conditions would generalize to less challenging missing rates. Moreover, we evaluate extreme missing settings during test (i.e., WSI-only and genomics-only), and our RMoE consistently achieves the best performance across all settings. – While our experiments consider such highly challenging missing settings, we will further investigate diverse missing rates for a more comprehensive robustness analysis. R3: HIGHLIGHTING CLINICAL MOTIVATION. – Thanks for recognizing the clinical relevance of our RMoE. Our method performs retrieval-based reasoning using clinically similar patients, which better aligns with real-world clinical workflows. We will emphasize this clinical motivation in the final version. R3: MNAR SETTING. – While this study focuses on standard MAR/MCAR benchmark settings, our retrieval-based completion strategy is potentially more robust to MNAR settings than purely generative approaches because it leverages biological similarity among clinically related patients. We will evaluate realistic MNAR settings where missingness correlates with clinical covariates in future work. R3: CLARIFICATION ON RETRIEVAL SPACE DESIGN. – We clarify that our design does not introduce optimistic bias or label leakage. Risk-group labels are used only as a coarse auxiliary signal to stabilize the retrieval process and organize clinically meaningful neighborhoods during training. Retrieved features provide contextual support for modality completion but never serve as surrogate labels.



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.

    Due to the mixed reviews, I recommend inviting the paper for rebuttal. The authors are encouraged to address the major concerns raised by the reviewers. If these issues can be adequately resolved during the rebuttal process, I would consider acceptance, even if the reviewers may not update their scores afterward.

  • 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 presents a retrieval-based mixture-of-experts framework for survival prediction with incomplete multimodal data. I agree with Reviewers #1 and #3 that the problem is clinically important, the proposed retrieval-based design is well motivated, and the experimental results are consistently strong across multiple settings. While Reviewer #2 raised concerns regarding the limited number of cohorts and missingness-rate evaluations, I view these as opportunities for future work rather than issues that undermine the main contribution. Overall, the paper provides a novel and practically relevant solution to multimodal missingness and I recommend acceptance.



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 reviewers agreed that the paper addresses an important and clinically realistic problem, namely survival prediction with incomplete multimodal data, and appreciated the retrieval-based “patient twin” formulation together with the reliability-aware MoE design. The method was viewed as technically sound, clinically well motivated, and empirically strong, particularly because it avoids potentially unreliable generative imputation and instead leverages patient-specific retrieval for missing modality handling. Although concerns remain regarding broader evaluation across additional cohorts, varying missingness settings, and non-MCAR scenarios, the post-rebuttal discussion overall supports the novelty, practical relevance, and potential impact of the work. I therefore recommend acceptance.



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.

    The paper tackles a clinically important incomplete multimodal survival prediction problem and proposes a reasonable retrieval-based MoE framework. Two reviewers support acceptance after rebuttal. Although evaluation is limited to three TCGA cohorts and one main missingness rate, the method is well motivated, technically sound, and shows consistent improvements. The final version should tone down broad robustness claims and discuss missingness assumptions and clinical-variable integration more clearly.



back to top