List of Papers Browse by Subject Areas Author List
Abstract
Multimodal survival analysis utilizing whole slide images (WSIs) and genomic profiles is fundamental for cancer prognosis. Recently, state-space models like Mamba have emerged as powerful tools for sequence modeling. However, translating this success to complex multimodal tasks is hindered by two critical limitations. First, conventional fusion strategies assume a static multimodal interaction strength, ignoring the fluctuating diagnostic importance of each modality across different patients and local regions. Second, the standard Mamba architecture processes tokens along predefined physical paths. This rigid scanning disrupts the semantic continuity of spatially scattered medical features and exacerbates long-range decay. To address these challenges, we introduce AdaSurvMamba as a novel adaptive framework for multimodal survival analysis. The framework features a Dual-Scale Importance-Aware Reconstruction (DSIR) module to dynamically modulate cross-modal interaction strength. It evaluates diagnostic importance at both the sequence and token levels to reconstruct the input representations. Furthermore, we propose a Semantic Aggregation Scanning (SAS) module to overcome contextual fragmentation. The SAS module dynamically reorganizes discrete tokens into semantically continuous sequences via a shared prototype pool. It explicitly modulates the state transition step size using global modality context and semantic priors to adaptively control the information absorption rate. Experiments across five TCGA cohorts demonstrate consistent gains over existing methods. Code is available at https://github.com/zjlGO/AdaSurvMamba.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/4611_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to the Code Repository
https://github.com/zjlGO/AdaSurvMamba
Link to the Dataset(s)
N/A
BibTex
@InProceedings{ZhoJia_AdaSurvMamba_MICCAI2026,
author = { Zhong, Jialong AND Liu, Tingwei AND Yue, Baokun AND Li, Jingjing AND Piao, Yongri AND Zhang, Miao AND Liu, Leiye AND Jiang, Jiahong AND Ji, Wei AND Lu, Huchuan},
title = { { AdaSurvMamba: Dynamic Fusion and Semantic Scanning for Multimodal Survival Analysis } },
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
5、This paper presents AdaSurvMamba, a multimodal survival analysis framework that integrates whole slide images (WSIs) and genomic profiles based on the Mamba state-space model. The work targets two key limitations of existing Mamba-based methods: static multimodal fusion that ignores varying modality importance, and rigid physical scanning that damages semantic continuity of scattered medical features. The main contributions are threefold: Propose the Dual-Scale Importance-Aware Reconstruction (DSIR) module to dynamically adjust cross-modal interaction strengths at both global sequence level and local token level. Design the Semantic Aggregation Scanning (SAS) module to reorganize tokens into semantically continuous sequences and adaptively control the state transition step size. Conduct extensive experiments on five TCGA datasets, showing that AdaSurvMamba outperforms multiple state-of-the-art multimodal survival prediction methods. The authors state that code will be released upon acceptance.
- 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 motivation is clear and clinically meaningful. The paper correctly identifies two critical issues in current Mamba-based multimodal survival analysis: static fusion and fixed scanning order, which are relevant challenges in medical multimodal learning. The technical framework is systematic and well-designed. The DSIR module realizes adaptive multimodal fusion, and the SAS module improves semantic continuity in state-space modeling. The overall architecture is logical and complete. Sufficient and rigorous experimental validation. The method is evaluated on five public TCGA cohorts, with comprehensive comparisons to recent SOTA models and detailed ablation studies on core modules, scanning strategies, and prototype settings. The paper is well organized and clearly written. The methodology is explained step by step, and the structure conforms to MICCAI conventions. The work targets cancer survival prediction, a high-impact task in computational pathology and precision oncology with good clinical potential.
- 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.
The innovation is relatively incremental. The proposed framework is developed based on existing Mamba-based survival methods (SurvMamba, ME-Mamba), and the ideas of adaptive fusion and semantic reorganization are similar to several recent visual Mamba improvements. Lack of external validation and clinical interpretability. All experiments are conducted on TCGA retrospective data, and no visualization or analysis of model attention is provided to enhance clinical trust. Ablation is only performed on two cohorts (BLCA, BRCA). More comprehensive validation across datasets would strengthen the conclusions. Some implementation details are insufficiently explained. Justifications for hyperparameters (e. g. , prototype numbers, loss weight) and comparisons of computational efficiency are missing. The paper does not discuss the model’s robustness to data noise or different feature extractors, which is important for real-world clinical applications.
- 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
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?
I recommend Weak Accept (4) because this paper addresses meaningful problems in multimodal survival analysis with a well-designed Mamba-based framework. The technical logic is sound, and the experimental results are consistent and convincing across five datasets. The work fits the scope of MICCAI and provides valuable insights for medical multimodal learning. Although the innovation is incremental and some clinical validation can be improved, the overall quality meets the acceptance standard of MICCAI.
- Reviewer confidence
Confident but not absolutely certain (3)
- [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.
Accept
- [Post rebuttal] Please justify your final decision from above.
The authors have provided a comprehensive and convincing rebuttal that addresses all the major concerns raised in my initial review. Specifically: They have clearly distinguished the novelty of AdaSurvMamba from prior Mamba-based survival methods, highlighting the unique design for heterogeneous WSI-genomic data. They have supplemented ablation results across all five TCGA cohorts, demonstrating the consistent effectiveness of the DSIR and SAS modules. They have provided additional details on hyperparameter selection, computational efficiency, noise robustness, and generalization across different feature extractors. They have committed to adding interpretability visualizations and clarifying implementation details in the revised manuscript. The authors’ responses are well-supported by existing experimental results and strictly adhere to the MICCAI rebuttal guidelines. The overall quality of the paper has been significantly strengthened, and I now fully support its acceptance.
Review #2
- Please describe the contribution of the paper
This paper presents a novel multimodal Mamba model for survival analysis. For survival prediction from pathological images and genomic profiles, the authors argue that existing multimodal Mamba-based approaches may be limited by fixed interaction strengths across modalities and fixed scanning paths, even though the factors most relevant to survival may differ by modality. Using TCGA datasets, they report improved performance of the proposed model in terms of C-index and provide ablation studies to support the importance of the proposed components.
- 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 paper addresses a clear and relevant multimodal survival-analysis problem, namely that different modalities may contain different survival-relevant factors. The proposed design is well motivated from this perspective. 2.The Methods section is generally well organized and easy to follow. 3.The ablation studies are useful and suggest that the proposed components contribute to the overall performance. In particular, the complementary effects of the two modules are 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.While the ablation studies suggest that the proposed modules are beneficial, the magnitude of improvement over the included baselines appears modest relative to the reported variation across the 5-fold cross-validation results. Overall, the current experiments support competitiveness more clearly than a robust performance advantage. 2.The empirical positioning with respect to prior Mamba-based survival models is not fully convincing. Although the paper cites related models such as SurvMamba and ME-Mamba, these methods are not included in the benchmark table. As a result, the distinction between the proposed contribution and prior Mamba-based survival approaches is not sufficiently established empirically. 3.This omission is particularly important because, under what appears to be a similar evaluation setting, namely the same TCGA datasets and 5-fold cross-validation, SurvMamba reports higher C-index values on BRCA and BLCA than those achieved by the proposed model. Although differences in backbone or implementation may affect direct comparability, these points should be discussed more explicitly.
- 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 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
1.The description of the genomic encoder is somewhat limited. Although there may be de facto standard choices for genomic encoding, additional details would help readers, especially as the genomic encoder in the proposed model is learnable. 2.The discussion of consistency with biological intuition regarding the optimal, or near-optimal, number of prototypes sounds somewhat stronger than what is currently supported. While pathological images are certainly heterogeneous, genomic profiles such as RNA-seq, CNV, and somatic mutation data may also be highly heterogeneous. A clearer rationale for the prototype-number sweep would strengthen this point.
- 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 relevant multimodal survival-analysis problem and proposes an interesting Mamba-based design. The manuscript is generally clear, and the ablation studies are helpful. However, the current empirical evidence does not yet convincingly establish a robust advantage over prior methods. In particular, the positioning relative to existing Mamba-based survival models is not fully clear, the benchmark set appears incomplete with respect to relevant competing methods, and the reported gains over the included baselines seem modest relative to the variation across folds. Overall, the idea is promising, but the present version leaves important questions unresolved.
- 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.
After reading the rebuttal, the reviewer views the paper more positively. The authors provided a reasonable explanation for the missing Mamba-based baselines and added a ME-Mamba comparison under a unified protocol, which substantially addresses one of the reviewer’s major concerns. The matched fold-wise statistical analysis also helps clarify that the gain is not only due to fold-level variation. The reviewer still considers the improvement modest and thinks the limitation regarding SurvMamba should be clearly stated. However, if the promised results and clarifications are included in the final version, the reviewer is willing to move the recommendation to Weak Accept.
Review #3
- Please describe the contribution of the paper
This paper proposes AdaSurvMamba, a multimodal Mamba-based survival analysis framework for jointly modeling WSI and genomic information. Specifically, the method introduces a Dual-Scale Importance-Aware Reconstruction (DSIR) module and a Semantic Aggregation Scanning (SAS) module to enhance multi-scale representation learning and semantic aggregation. Experiments on five public TCGA cohorts demonstrate favorable performance against multiple competing methods, while the ablation studies further validate the effectiveness of the proposed modules from both global and local perspectives.
- 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 proposed DSIR and SAS modules are reasonably novel, and the method section provides relatively complete and clear mathematical formulations. 2 The experimental setup, training details, and implementation pipeline are described in sufficient detail, making the work fairly reproducible. 3 The comparative experiments and ablation studies are well designed and provide reasonable support for both the overall superiority of the method and the effectiveness of the proposed modules at different levels. 4 The paper is well organized overall, with a clear technical pipeline, well-presented figures and tables, and a generally coherent narrative.
- 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 Although the paper already includes comparisons with a broad range of related multimodal survival analysis methods, the naming of the proposed method as AdaSurvMamba naturally leads readers to associate it with the existing SurvMamba. Therefore, a direct empirical comparison would further help clarify the contribution of the present work. 2 In Figure 1, within the Semantic prototype construction part of Semantic Aggregation Scanning, the symbol Q_h appears to be a likely typo; based on the context, it may instead be intended as Q_g. The authors are encouraged to verify this notation and ensure consistency throughout the manuscript.
- 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 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
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?
This paper presents a fairly complete method design for multimodal survival analysis, with reasonably novel modules and sufficiently thorough experimental and ablation studies. Overall, the paper forms a coherent line of argument that generally supports its main conclusions. At the same time, the manuscript could better clarify its relationship to highly related prior work, and there are also a few minor notation-level issues that should be corrected.
- Reviewer confidence
Confident but not absolutely certain (3)
- [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.
Accept
- [Post rebuttal] Please justify your final decision from above.
The authors adequately addressed the main concerns raised in the initial review. In particular, they provided a reasonable and sufficiently detailed explanation for why SurvMamba could not be reliably reproduced.
Author Feedback
To R1: We appreciate your positive assessment. W1: Novelty. Our method is not a direct extension of prior Mamba-based survival/visual models. SurvMamba/ME-Mamba use fixed cross-modal interaction or predefined/importance-based scans, whereas DSIR/SAS learn patient-/token-level modality relevance and semantic continuity for adaptive weighting and semantic reordering before scanning. Recent Mamba methods (e.g., [6,12,27]) improve scanning/fusion via spatial/2D paths or fixed/shared gated fusion for regular spatial data. In contrast, WSI-genomics survival has sparse/scattered prognostic patches and unordered genomic features; thus our design targets this heterogeneous setting rather than reusing prior Mamba designs. W2: Thanks for the suggestion. We have visualized WSI patch-level attention heatmaps and modality-importance maps. Due to rebuttal format limit, we will include and discuss them in the revision. External validation. We agree external validation is important. However, external cohorts with matched WSI, genomic profiles, and survival outcomes are difficult and costly to obtain. Following [21,22,24,25], we evaluate on five TCGA cohorts in Table 1 to assess generalization across cancer types. W3: Tabs. 2-4 ablate key designs on BLCA/BRCA, showing gains beyond one cohort. We further conducted ablations on all five cohorts; on average, DSIR and SAS improve the baseline by 4.46% and 3.25%. Due to space limits, we will include averaged results in the revision and per-cohort details on GitHub. W4:Thank you for the suggestion. Prototype-number analysis is in Table 5.For the loss weight, λ = 0.5 performs best among {0.2, 0.5, 1.0}; λ=0.2/1.0 reduce the average C-Index by 0.4%/0.9%. This shows that too weak/strong modality-specific supervision can hurt multimodal optimization. For efficiency, under 1000 WSI patches, ours uses 1.28 GFLOPs/3.46M parameters, vs. 1.53GFLOPs/3.83M for ME-Mamba. We will add these details to the revision. W5: We test noise robustness by adding Gaussian noise ϵ∼N(0,0.01) to WSI/genomic features at test time. Our method and the second-best show C-Index drops of 1.74% and 1.82%. With perturbation training, our drop decreases to 1.06%. We will discuss this to encourage future research. Besides ResNet50/UNI in Table 1, we further test CONCH, where our model still outperforms the second-best by 2.56% on average.
To R2: We thank you for the constructive comments and will address each concern. W1: To assess reliability, we compared matched fold-wise C-index scores between our model and the second-best under identical 5-fold splits using a paired t-test. The gain is significant (p<0.05) and we will add this analysis. W2/W3: We would like to clarify this point. ME-Mamba was accepted only a few days before the MICCAI deadline, and its code was unavailable, making inclusion difficult. After submission, we contacted the authors, obtained the code, and reproduced it under our unified protocol. Our model outperforms ME-Mamba on all five cohorts (+2.53%/+2.64% avg. C-index under ResNet50/UNI), which will be added to the benchmark table. For SurvMamba, we also tried to obtain the official implementation, but the code remains unavailable. As noted, reported numbers may differ due to different settings; thus, we cannot reliably reproduce it under the same protocol. We will discuss this limitation and clarify the distinction from prior Mamba-based survival models (please refer R1-W1). Opt: We will add more details on the genomic encoder. We agree that genomic data can be heterogeneous. A smaller number of genomic prototypes encourages pathway-level summarization of RNA-seq, CNV, and mutation signals, captures major molecular programs, and avoids over-fragmentation. We will discuss it in detail in the revision.
To R3: We appreciate your positive comment. W1: Thank you for your suggestion. Please refer to R2-W2/W3 for details. W2: We apologize for this typo and will correct it while checking notation consistency.
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 proposes a well-motivated method AdaSurvMamba to address the issues of static multimodal fusion and semantically inconsistent sequence modeling in Mamba. It introduces DSIR and SAS modules to enable adaptive fusion and semantically structured sequence modeling. However, there are concerns regarding the experimental validation, particularly the lack of strong baseline comparisons.
- 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 concerns regarding the paper have been addressed, and all reviewers agree to accept the paper.
Meta-review #2
- After you have reviewed the rebuttal and updated reviews, please provide your recommendation based on all reviews and the authors’ rebuttal.
Accept
- Please justify your recommendation.
This paper proposes a multimodal Mamba-based survival analysis framework with adaptive fusion and semantic scanning modules. Reviewers appreciate the clear motivation, coherent method design, and empirical evaluation across multiple TCGA cohorts. After rebuttal, most of the concerns are reasonably addressed. Some limitations remain regarding external validation and clinical interpretability, which should also be addressed in revision.
Meta-review #3
- After you have reviewed the rebuttal and updated reviews, please provide your recommendation based on all reviews and the authors’ rebuttal.
Accept
- Please justify your recommendation.
The paper received mixed reviews and it was invited for rebuttal. The authors submitted a rebuttal to address the raised points of the reviewers. After the rebuttal period all the reviewers indicated that their concerns had been properly addressed in the rebuttal and kept a positive score. The AC read the paper, reviews, rebuttal and final justification and agrees that the paper has merit and it would be interesting to be presented in the community.