List of Papers Browse by Subject Areas Author List
Abstract
Recent multimodal survival prediction models have demonstrated strong predictive performance by leveraging complementary information across modalities. However, such models generally assume data completeness and exhibit limited robustness toward missing modalities, which are frequently encountered in real-world clinical settings. We propose the Evidential Missing Modality Survival Fusion model for multimodal survival prediction under missing modalities. EMMS offers a straightforward, computationally effective approach to survival analysis without requiring a generative phase for missing data. By employing Dempster–Shafer theory and the newly introduced Gaussian Random Fuzzy Numbers for multimodal decision fusion, it considers both aleatoric and epistemic uncertainty alongside modality reliability for fusion. Moreover, the model treats missing modalities as vacuous evidence, preventing interference with available inputs and naturally reflecting increased uncertainty and calibrated predictions. Extensive experiments on four cancer datasets demonstrate state-of-the-art performance while providing calibrated and interpretable uncertainty estimates under incomplete multimodal observations, without extra computational overhead. Code is available at https://github.com/YuchengXing99/EMMS.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/6476_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to the Code Repository
https://github.com/YuchengXing99/EMMS
Link to the Dataset(s)
N/A
BibTex
@InProceedings{XinYuc_Evidential_MICCAI2026,
author = { Xing, Yucheng AND Mo, Hailan AND Wang, Zi AND Huang, Ling AND Feng, Mengling},
title = { { Evidential Fusion Network for Multimodal Survival Prediction Under Missing Modalities } },
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
This paper proposes an Evidential Missing Modality Survival Fusion (EMMS) framework to address the common issue of missing modalities in multimodal cancer survival analysis. The method leverages Dempster–Shafer theory and Gaussian Random Fuzzy Numbers (GRFN) to model uncertainty and treats missing modalities as vacuous evidence, enabling an imputation-free fusion strategy. Demonstrating the method on multiple TCGA cancer datasets under different missing-modality scenarios.
- 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.Addressing missing modalities in multimodal survival analysis is important for real-world deployment. 2.The paper evaluates different missing rates, which improves the practical relevance of the study. 3.Experiments are conducted on four TCGA cancer cohorts, which provides a relatively diverse validation setting.
- 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.Evidential Deep Learning (EDL) and GRFN-based survival modeling are not new (see [17, 18]). The main contribution of this work is applying these existing techniques to the missing modality setting, which is incremental and does not constitute a substantial methodological advance. 2.Some abbreviations such as ERFS are not defined at first use, naming is inconsistent (e. g. , Dispro vs. DisPro), and inconsistent bold formatting applied to both full sentences and individual words. 3.The symmetric KL divergence-based Alignment Loss (Eq. 11) for aligning pathology and genomics predictions is problematic, given that forcing distributional similarity between inherently heterogeneous modalities risks compromising complementary information, particularly in cases of modality conflict. The Prototype Learning component (Eq. 4) lacks specification of the prototype number K and initialization strategy, which severely hinders reproducibility and prevents a rigorous evaluation of the method’s robustness. 4.The experimental comparison is limited. The paper does not include several important missing-modality methods such as SMIL [20], M3Care [28], and Flex-MoE [27]. Using only DisPro and MUSE as baselines is insufficient to support claims of strong or state-of-the-art performance.
- 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.
(2) Reject — should be rejected, independent of rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
1.Methodologically, it lacks substantial novelty, as it only incrementally applies existing EDL and GRFN techniques to the missing modality setting. 2.The proposed Alignment Loss is problematic, risking loss of complementary information between heterogeneous modalities, and critical details of Prototype Learning are unspecified, undermining reproducibility. 3.The experimental evaluation is incomplete: key missing-modality baselines are excluded, and only two baselines are insufficient to support state-of-the-art performance claims. 4.Persistent presentation issues (undefined abbreviations, inconsistent naming/formatting) further reduce the manuscript’s professionalism and readability.
- 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.
Although the rebuttal provides some clarifications, I still find the methodological novelty insufficient, the method comparison inadequate to support the SOTA claim, and several key implementation details insufficiently specified.
Review #2
- Please describe the contribution of the paper
This paper introduces the EMMS model for multimodal survival prediction with missing modalities. It uses Dempster-Shafer theory and GRFN to fuse two modalities. It treat missing modalities as vacuous evidence with zero epistemic strength. Experiments on four TCGA datasets show satisfying performance under different modality missing 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.It is a novel idea to use the Dempster–Shaferevidential theory and GRFNs to handle the modality missing problem, bypass the imputation hallucination and imputation-free’s information loss.
2.The paper is clearly written, well-structured 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.How many prototypes are used in each branch? It is also unclear whether these prototypes are semantically meaningful, for example, whether they correspond to distinct risk patterns. In addition, how are the prototypes initialized? Since there appears to be no explicit mechanism to encourage diversity or separation among them, what prevents the prototypes from becoming redundant or collapsing during training?
2.The interaction between modalities appears to be rather weak, as it is introduced only at the final Gaussian-distribution level. Since the WSI and genomic features are encoded independently and no explicit cross-modal feature interaction is modeled, could this lead to inconsistent or divergent modality-specific predictions before fusion?
3.Why are the epistemic strengths produced by the two unimodal branches directly comparable? Since the pathology and genomics branches are trained separately and may have different scaling behaviors, it is unclear what guarantees that their epistemic strengths are calibrated onto a common scale and can therefore be reliably used for cross-modal weighting in fusion.
- 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?
Despite some weaknesses that still need to be clarified, this paper proposes a novel strategy for handling missing modalities in survival modeling that is both efficient and effective. Overall, I recommend accepting this paper.
- 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 are solved basically.
Review #3
- Please describe the contribution of the paper
The authors propose a method for missing data estimation based on uncertainty theory (Dempster-Shafer theory) and Gaussian random fuzzy numbers. Estimating based on a population is certainly an interesting approach for estimating and comparing with existing imputation approaches. Authors compare performance for no-missing modalities and 60% missing data. Their performance (c-index) has reduced, as expected. It would be nice to compare against a couple of known approaches and a random or neighborhood-based) approach. The results presented in Tables 1 and 2 are hard to follow, especially regarding what is being compared. It would help to report performance statistics in the abstract.
- 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.
propose a method for missing data estimation based on uncertainty theory (Dempster-Shafer)
- 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 reviews. Present key results in the abstact.
- 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?
methods development
- Reviewer confidence
Confident but not absolutely certain (3)
- [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.
N/A
- [Post rebuttal] Please justify your final decision from above.
N/A
Author Feedback
We thank all reviewers for their constructive comments. We are encouraged that R2 (Accept) and R3 (Weak Accept) recognize the novelty and efficiency of our framework and primarily address R1’s concerns below.
[R1] Novelty. We acknowledge that the unimodal survival modeling component of EMMS is related to prior GRFN-based methods. However, our main contribution is the treatment of missing modalities, which these works do not address. We introduce vacuous evidence as a DS-theoretic formulation for missing modalities within the GRFN framework, yielding a closed-form fusion rule (Eq. 9) that naturally falls back to the available modality when another is absent without imputation or heuristic masking. The formulation also provides an interpretable uncertainty term h_f reflecting epistemic uncertainty from missing information. As R2 explicitly notes, this bypasses both imputation hallucination and the information loss common in imputation-free approaches. To our knowledge, this is the first work combining DS theory and GRFNs for missing-modality survival prediction.
[R1] Typos. “ERFS” = Epistemic Random Fuzzy Sets; “DisPro” capitalization and bold formatting inconsistencies have been corrected throughout.
[R1] Alignment loss vs. complementarity. We respectfully note a misinterpretation. L_align operates at the decision level, not on feature representations z_p and z_g. Modality-specific information is therefore fully preserved upstream. A stop-gradient on variance ensures gradients from L_align affect only mu_p and mu_g, while sigma_p^2, sigma_g^2, h_p, and h_g remain unchanged, preserving each modality’s uncertainty profile and evidential strength. Since both modalities supervise the same event time t, aligning mu_p and mu_g regularizes central tendency only, not distributional shape. Residual disagreement is absorbed into sigma_f^2 by DS fusion as calibrated uncertainty.
[R1, R2] Prototype details. K=50 prototypes per branch are deterministically initialized via k-means on training features, which partitions the feature space into well-separated regions and ensures initial diversity. During training, RBF affinity induces competitive activation among prototypes, and the product-intersection (Eq. 7) naturally down-weights prototypes with low affinity scores, preventing collapse without requiring an explicit diversity loss.
[R1, R3] More baselines. SMIL [20], M3Care [28], and Flex-MoE [27] are designed for general classification tasks and were not considered survival prediction baselines. Nevertheless, we thank the reviewers for mentioning them and have 1) adapted Flex-MoE [27] for the survival task and 2) implemented a KNN imputer as suggested by R3 for comparison. EMMS consistently achieved the best performance across all four datasets. EMMS outperforms Flex-MoE by +0.037/−0.014 in mean C-index/IBS under 0% missing and by +0.069/−0.013 under 60% missing, and outperforms KNN by +0.103/−0.018 under 60% missing. Full results have been included in the final version.
[R2] Divergent predictions. During training, L_align reduces inter-branch divergence by encouraging consistent predictions for the same target y=log(T). At inference, DS fusion (Eq. 9) handles remaining disagreement: the branch with higher epistemic strength h contributes more to mu_f, while disagreement is reflected in sigma_f^2, supported by strong calibration results (Fig. 2).
[R2] Comparability of h. Prototype precisions h_k are initialized identically across both branches. The fused survival loss Lf_surv (Eq. 10) jointly backpropagates into h_p and h_g, providing a shared supervisory signal that continuously aligns their epistemic strengths during training.
[R3] Readability. We have (i) included headline C-index and IBS in the abstract, (ii) reorganized Tables 1-2 with clearer G-only/P-only/Complete grouping, and (iii) revised captions accordingly.
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 presents a well-motivated, uncertainty-aware method for cancer survival prediction under missing-data situations. The reviewers raised critical concerns regarding technical details, experimental justifications, and methodological novelty.
- 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 authors’ rebuttal and clarifications addressed most concerns.
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 proposes EMMS, an Evidential Missing Modality Survival Fusion framework for addressing missing modalities in multimodal cancer survival analysis. Its focus on missing modalities is both practical and clinically important. The reviewers raised several important concerns, but the authors addressed them comprehensively in the rebuttal, including adding baselines such as Flex-MOE and further demonstrating the effectiveness of the proposed method. Overall, the work is solid and relevant, and I 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.
Two out of three reviewers recommend acceptance after the rebuttal. The application of Dempster-Shafer theory and Gaussian Random Fuzzy Numbers to missing-modality survival prediction is a technically sound contribution.
R2’s concerns were satisfactorily addressed. While R1’s concerns about limited novelty and insufficient baselines are not fully resolved, the rebuttal clarified the key misreading of the alignment loss and provided additional baseline comparisons.
The authors are required to address the following in the camera-ready version: 1) fix all presentation issues (undefined abbreviations, inconsistent naming, etc) 2) specify prototype number K and initialization strategy 3) clarify the comparability of epistemic strengths across branches 4) improve the readability of Tables 1+ 2