Abstract

In real-world clinical settings, multimodal medical data are often characterized by severe modality imbalance and structured missingness. In our tri-modal lung cancer cohort (WSI, CT, RNA), 75.4% of patients lack CT scans. Most existing multimodal learning methods implicitly assume complete modality availability or rely on naive imputation, leading to catastrophic performance collapse when applied to highly incomplete deployment data. To address this challenge, we propose RAM-Missing, a Retrieval-Augmented Missing-aware fusion framework designed for robust lung cancer subtyping under severe missingness. Specifically, our framework introduces three key components: (1) We design a Retrieval-Augmented Proxy Generator that constructs proxy embeddings for missing CTs by querying a historical memory bank. (2) We introduce a Retrieval Uncertainty Quantifier that mathematically estimates retrieval entropy to gauge the reliability of these proxies. (3) We propose an Uncertainty-Guided Conditional Router that dynamically allocates fusion experts based on the modality mask and uncertainty signal, effectively preventing the model from over-trusting low-quality imputed features. Extensive experiments using patient-level 5-fold cross-validation demonstrate that RAM-Missing consistently mitigates catastrophic degradation under extreme missing ratios, achieving superior prevalence-weighted AUROC and cross-fold stability compared to conventional fusion strategies. The source code is publicly available at https://github.com/Flliin/RAMMissing.

Links to Paper and Supplementary Materials

Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/5091_paper.pdf

SharedIt Link: Not yet available

SpringerLink (DOI): Not yet available

Supplementary Material: Not Submitted

Link to the Code Repository

https://github.com/Flliin/RAMMissing

Link to the Dataset(s)

N/A

BibTex

@InProceedings{LiFul_RAMMissing_MICCAI2026,
        author = { Li, Fulin AND Cai, Qing AND Zhang, Shugang AND Li, Jinxing},
        title = { { RAM-Missing: Retrieval-Augmented Missing-Aware Fusion for Robust Lung Cancer Subtyping } },
        booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
        year = {2026},
        publisher = {Springer Nature Switzerland},
        volume = {LNCS 16886},
        month = {September},
        page = {pending}
}


Reviews

Review #1

  • Please describe the contribution of the paper

    This work presents a retrieval-augmented framework for dealing with missing modality in lung cancer subtyping. Specifically, the authors consider the missing of CT scans while WSI and RNA are available. Missing CT is dealt with embedding-based cross-modality retrieval, and then the three modalities are fused by an uncertainty-guided conditional router. The method was validated on the TCGA lung cancer dataset, showing that the proposed method could achieve robust performance under high missing ratios.

  • 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 authors deal with a practical problem in cancer subtyping, where missing modality affect the performance of AI systems 2, the design of MoE fusion can deal with multi-modality and fuse them adaptively 3, The method show stable performance under different CT missing ratio

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

    1, The authors assume that Eq. 1 can retrieval similar cross modality cases. However, this assumption may not hold when the three encoders are not aligned during pretraining. It seems that the three encoders are pretrained separately (WSI encoder is ResNet50 pretrained on ImageNet, while 3D ResNet is trained on CT). As they are not well aligned, one cannot ensure that the feature similarity-based retrieval works. Therefore, the rational of the method is limited. In addition, the authors did not report the performance of cross-modality retrieval, so it is unclear how this step is reliable.

    2, The missing modality problem setting does not match clinical practice. The author said that “High-cost imaging modalities such as CT are often available only for a subset of patients, whereas pathology and transcriptomic data are more routinely collected”. This is not correct in most clinical environment. CT scanning is much cheaper than WSI and transcriptomic data. For lung cancer patient, CT scanning is usually first obtained for preliminary diagnosis, if the patient is suspicious of cancer, then WSI is obtained for further diagnosis. Especially in underdeveloped regions, patients may not have WSI and RNA information, but CT scans are more available. It is more practical to consider the missing of WSI/RNA rather than the missing of CT.

    3, The method was only validated on one dataset, which makes it generalizability to other datasets unknown. The author also did not consider the missing of WSI and RNA.

    4, The experimental results are unconvincing. Table 1 shows that only using WSI or RNA leads to an AUROC of 0.68 and 0.64, respectively. When CT is missing, a baseline method combining WSI and RNA is expected to achieve AUROC higher than 0.68.However, the authors’ method only has an AUROC of 0.509, which is much lower than the baseline and single modality prediction.

  • 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.

    (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?

    The problem setting does not match clinical practice, and the method rationale is weak, with unconvincing experimental results

  • 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.

    Due to the unrealistic problem setting, poor alignment between the encoders of different modalities and unconvincing results, I keep my score for the review. Though the authors tried to explain these issues, the reply does address the concerns.



Review #2

  • Please describe the contribution of the paper

    This paper proposes RAM-Missing, a retrieval-augmented missing-aware fusion framework for multimodal lung cancer subtyping under severe modality imbalance. The method addresses a realistic clinical setting where CT scans are frequently missing by constructing proxy CT embeddings through a memory-based retrieval mechanism. It further introduces an entropy-based uncertainty quantification to estimate the reliability of retrieved proxies and integrates this signal into a mixture-of-experts (MoE) fusion module via a conditional routing mechanism. The framework is evaluated on a tri-modal TCGA cohort (WSI, CT, RNA) and demonstrates improved robustness and stability under high missingness scenarios compared to standard multimodal fusion approaches.

  • 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.

    Clinically relevant problem formulation: The paper explicitly addresses severe modality missingness (e.g., 75.4% CT missing), which is highly realistic in clinical practice and often overlooked in multimodal learning .

    Well-structured and interpretable framework: The proposed architecture (Fig. 1) clearly separates modality encoding, retrieval-based proxy generation, uncertainty estimation, and adaptive fusion, making the design easy to follow and interpret .

    Uncertainty-aware fusion mechanism: Incorporating entropy-based uncertainty into the routing decision is a meaningful contribution that helps mitigate over-reliance on potentially noisy imputed modalities.

    Demonstration of catastrophic failure in baselines: The paper provides strong evidence that standard fusion methods collapse under missing modalities (Table 1), which is an important and valuable insight.

    Comprehensive ablation studies: The contribution of each module (retrieval, router, uncertainty) is systematically evaluated (Table 2), strengthening the validity of the design.

    Robustness analysis: Evaluation across varying missingness levels (Table 3) provides useful insight into model stability under realistic deployment conditions.

    Use of prevalence-weighted metrics: The adoption of weighted AUC aligned with real-world modality distribution is appropriate and improves evaluation realism.

  • 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.

    Outdated WSI backbone choice: The WSI encoder relies on a ResNet-50 pretrained on ImageNet with global average pooling (p.3) . However, recent work in computational pathology has demonstrated the superiority of self-supervised and transformer-based models (e.g., DINO/ViT variants) for capturing fine-grained morphological features. The authors do not justify this design choice, which may limit performance.

    Suboptimal pooling strategy: The use of global average pooling for slide-level aggregation may dilute discriminative signals in heterogeneous WSIs. Alternative approaches such as max pooling or attention-based pooling are often more effective but are not explored or discussed.

    Retrieval proxy validity is not validated: The framework critically depends on retrieval-based proxy CT embeddings (p.4) , yet there is no direct evaluation of whether these proxies meaningfully approximate true CT features. This raises uncertainty about whether performance gains stem from meaningful reconstruction or architectural regularization.

    Limited performance gains: The improvement in weighted AUC (e.g., ~0.546 to 0.564 in Table 1) is relatively modest , and the clinical significance of this gain is not clearly justified.

    Incomplete evaluation metrics: The paper reports AUC-based metrics but does not include F1-score or other threshold-dependent metrics, which are important for evaluating classification performance in imbalanced clinical settings.

    Limited dataset diversity: The evaluation is restricted to a single TCGA cohort without external validation, raising concerns about generalizability.

    No clinical/expert validation: There is no assessment of whether retrieved proxies or predictions align with clinically meaningful patterns.

    Simplistic uncertainty estimation: The entropy-based uncertainty formulation may not fully capture the reliability of retrieved proxies and lacks deeper validation.

  • 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 authors claimed to release the source code and/or dataset upon acceptance of the submission.

  • Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?

    N/A

  • Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2026/en/REVIEWER-GUIDELINES.html

    The paper addresses an important and practical challenge in multimodal medical AI. To further strengthen the work, the authors are encouraged to:

    Justify the choice of ResNet-50 as the WSI backbone, or compare with modern self-supervised/pathology-specific encoders (e.g., DINO-based models).

    Explore alternative feature aggregation strategies (e.g., max pooling or attention-based pooling).

    Provide additional evaluation metrics such as F1-score for better clinical interpretability.

    Validate the quality of retrieval-based proxy embeddings (e.g., similarity to ground-truth CT features).

    Clarify the practical significance of the observed performance gains.

    Provide external validation or cross-dataset evaluation.

    Include clinical or expert-based assessment if possible.

    Addressing these points would significantly strengthen the contribution.

  • 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?

    This paper addresses a highly relevant problem in multimodal medical learning—robust performance under severe modality missingness—and proposes a well-motivated retrieval-augmented fusion framework with uncertainty-aware routing. The approach is technically sound and demonstrates improved robustness compared to standard baselines, particularly in preventing catastrophic performance degradation.

    However, several limitations reduce the overall impact. The method relies on relatively simple unimodal encoders (e.g., ResNet-50 for WSI) and basic pooling strategies, which may not fully exploit available data. The evaluation lacks important metrics such as F1-score, and the observed improvements in AUC are relatively modest, raising questions about practical significance. Additionally, the key component of retrieval-based proxy generation is not sufficiently validated.

    Overall, the paper presents a solid contribution with meaningful insights but would benefit from stronger validation, improved modeling choices, and more comprehensive evaluation.

  • Reviewer confidence

    Very confident (4)

  • [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



Review #3

  • Please describe the contribution of the paper

    This paper proposes RAM-Missing, a retrieval-augmented missing-aware fusion framework for multimodal lung cancer subtyping (LUAD vs. LUSC) under severe structured modality missingness. The setting involves tri-modal data (WSI, CT, RNA), where CT is missing in 75.4% of patients, reflecting a realistic and clinically relevant deployment scenario.

    The method introduces three key components: (1) a retrieval-based CT proxy generator using a memory bank of complete cases, (2) an entropy-based uncertainty estimator for retrieval reliability, and (3) a conditional Mixture-of-Experts (MoE) fusion mechanism that dynamically routes information based on modality availability and uncertainty signals. Extensive experiments on TCGA with patient-level cross-validation and realistic missingness patterns demonstrate strong robustness and significant mitigation of catastrophic performance degradation.

  • 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.Strong clinical relevance and realistic problem setting The paper addresses an important and practically relevant problem: multimodal cancer diagnosis under severe and structured missingness. The reported CT missing rate (75.4%) reflects real-world clinical constraints rather than artificially simulated missing patterns, significantly increasing the practical value of the study. 2.Well-motivated robustness objective Unlike many multimodal learning works that focus solely on average-case performance, this paper explicitly targets worst-case robustness and failure prevention under missing modalities. The use of prevalence-weighted AUROC and stress testing under varying missing ratios is appropriate and clinically meaningful. 3.Clear and coherent methodological design The proposed framework is well-structured and logically consistent: Retrieval-based proxy construction effectively addresses missing CT modalities. Uncertainty estimation via entropy provides a principled signal for reliability assessment. Conditional MoE routing enables adaptive fusion and mitigates the risk of relying on noisy imputed features. The overall architecture is well-motivated and easy to follow. 4.Strong robustness and stability analysis The experimental results demonstrate consistent improvements in missing-CT scenarios and, more importantly, significant reduction in catastrophic performance collapse (AUC < 0.5 cases). The stress test under increasing missingness ratios further strengthens the claim of robustness. 5.Potential clinical impact The emphasis on avoiding failure cases rather than solely improving average performance is particularly important for medical applications. The proposed uncertainty-aware routing mechanism improves interpretability and may increase clinician trust in multimodal AI systems.

  • 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.Moderate methodological novelty While the overall framework is well-designed, the individual components (retrieval-based imputation, entropy-based uncertainty, and MoE routing) are relatively well-established. The novelty lies primarily in their integration and application to a realistic missing-modality clinical setting. However, given the strong motivation and careful system design, this does not significantly reduce the contribution. 2.Limited external validation The evaluation is restricted to TCGA data. Although the missingness simulation is realistic, cross-institutional validation would further strengthen claims regarding generalizability. 3.Heuristic nature of uncertainty modeling The uncertainty estimation is based on Shannon entropy of retrieval weights, which is intuitive but relatively simple. A more principled probabilistic calibration analysis could further improve methodological rigor.

  • 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.

    (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?

    This paper addresses a clinically important and underexplored problem of multimodal learning under severe missingness, providing a well-designed and empirically validated framework that significantly improves robustness and failure prevention in realistic deployment settings.

  • 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 the reviewers for the constructive comments. Reviewer #1 Q1: Retrieval validity. A1: Although WSI, RNA and CT use different encoders, they describe the same patient/case/clinical state, giving weak patient-level consistency. Retrieval uses top-k=5, not a single top-1 neighbor, and does not claim CT reconstruction. Train-fold-only audit: top-1 label agreement 67.6%, top-5 any-hit 81.6%; the proxy is phenotype-level evidence guarded by conservative fallback. Q2/Q3: Missing-CT setting and single-dataset generalization. A2: Missing-CT is not artificial, but structural missingness in a patient-level aligned tri-modal lung cancer cohort. Tri-modal modeling is clinically motivated: WSI captures morphology/microenvironment, RNA captures molecular programs, and CT captures radiographic phenotype. Open WSI+RNA+CT cohorts remain scarce due to clinical data-sharing and patient-level alignment constraints. The audited CPTAC subset lacks CT, so it cannot support full tri-modal external validation; this is stated as limitation/future work. Q4: Experimental comparison. A4: Single-modality models are trained for one modality, whereas missing-modality evaluation removes a key modality from the same multimodal framework to test robustness. The proper Missing-CT baseline is the observed-modality WSI+RNA pathway and methods under the same CT-missing condition. Reviewer #2 Q1: WSI encoder. A1: We compared UNI, ViT-B/16 and DINO ViT-S/16.Strong encoders improve the WSI ceiling, but under strong baselines gains are about 1-2%, while extraction cost and model size increase substantially. DINO ViT-S/16 has about 21.7M parameters, whereas the RAM-Missing-CF fusion head has about 0.38M trainable parameters, 57x smaller; DINO offline extraction is about 9x the lightweight fusion-head training time. Our focus is small-parameter missing-aware fusion on frozen features. Q2: GAP/MIL pooling. A2: We ran pooling ablations, including MeanPool+LogReg, MaxPool+LogReg and AttentionMIL. The current feature/pooling strategy was more stable, so we kept it as the main setting and will add this ablation in the final version or supplement. Q3: Proxy validity. A3: See R1-Q1.The proxy is top-k phenotype-level evidence, not exact CT reconstruction; stronger retrieval learning and explicit cross-modal alignment are future work. Q4: Limited gain. A4: Value is not only AUROC gain, but also robustness, trainable-parameter efficiency and deployment cost. Our small fusion head avoids end-to-end retraining of a large WSI encoder, lowering training/deployment cost under scarce tri-modal samples. CVPR 2023 Multimodal Prompting With Missing Modalities similarly uses few learnable parameters instead of heavy retraining. Q5: Metrics. A5: We will add AUPRC, Macro-F1 and Balanced Accuracy. Q6: Dataset diversity. A6: See R1-Q2/Q3.Full tri-modal external validation is limited by CT availability and patient-level alignment. Q7: Expert validation. A7: We will report clinician review of test cases/retrieved neighbors where space permits, connecting retrieved evidence with clinical patterns. Q8: Entropy uncertainty. A8: Entropy is a retrieval reliability signal, not calibrated uncertainty. Concentrated top-k weights indicate consistent proxy evidence; diffuse weights indicate higher proxy risk. Normalized entropy H=-sum_j alpha_j log(alpha_j)/logK triggers conservative routing toward observed WSI/RNA evidence. Reviewer #3 Q1: Novelty. A1: Prior work studied dual-modality fusion or isolated missing-modality cases. Our contribution is an integrated workflow for structurally incomplete WSI/RNA/CT lung cancer subtyping: retrieval-informed evidence, missing-aware routing, conservative fallback and fixed patient-level evaluation for scarce tri-modal clinical cohorts. Q2/Q3: External validation and uncertainty. A2: See R1-Q2/Q3 and R2-Q8: full tri-modal external validation is limited by CT availability; entropy is a routing risk signal rather than calibrated uncertainty.




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 a clinically relevant problem: multimodal lung cancer subtyping under severe modality missingness. Reviewers agreed that handling missing data robustly is important. The proposed retrieval-augmented fusion framework with uncertainty-aware routing was considered well-motivated.

    At the same time, there are significant concerns that should be addressed in rebuttal. Reviewer #1 questioned the clinical realism of assuming CT is the missing modality rather than WSI or RNA, and also challenged the validity of the retrieval mechanism, noting that cross-modal encoder alignment and retrieval quality were not directly validated. Reviewer #1 also highlighted unconvincing results where missing-modality performance appears worse than simpler baselines.

    Additional concerns that can’t be fully resolved in rebuttal include evaluation on only a single TCGA cohort, modest performance gains, and limited evidence that retrieved CT proxies meaningfully approximate true CT information.

  • After you have reviewed the rebuttal and updated reviews, please provide your recommendation based on all reviews and the authors’ rebuttal.

    Reject

  • Please justify your recommendation.

    The main issue is the unconvincing empirical result under the main deployment scenario: when CT is missing, the proposed method achieves only 0.51 AUROC, which is worse than using WSI or RNA alone (0.68 and 0.64). This directly contradicts the paper’s main claim and was not convincingly addressed in the rebuttal.

    Moreover, the retrieval mechanism, the core technical contribution, relies on cross-modal similarity between independently pretrained encoders that are not aligned. Since no direct validation of retrieval quality is provided, it’s hard to evaluate how cross-modal similarity between unaligned encoders could help. The authors acknowledge this limitation.



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 majority of the reviewers provided positive recommendations. I agree with the reviewers that the primary remaining concern relates to the problem setting. However, the paper demonstrates reasonable scientific merit, and the multimodal application to missing-modality learning in lung cancer represents a relevant and relatively novel application area. Therefore, I lean toward a positive decision.



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.

    Overall, the paper is well motivated and the proposed method is technically sound. A critical concern regarding the clinical realism of the methodological design hampers the strength of the paper.



back to top