Abstract

Multi-modal survival analysis enables patient risk stratification and clinical decision-making in oncological prognosis. However, existing methods often suffer from heterogeneous domain shifts in real-world cross-site deployment, which refer to data discrepancies from inter-center equipment and modality-specific protocol variations. We observe that survival ordinal characteristics remain invariant across clinical sites and modalities. Leveraging this insight, we propose SurvTTA, the first test-time adaptation framework for accurate multi-modal survival analysis under heterogeneous domain shifts. SurvTTA bridges multi-modal domain gaps by maintaining reliable topological risk-order consistency and modality-wise aligned representation, enabling stable adaptation of pretrained models across different clinical sites with only unlabeled data. Specifically, a novel order-aware regularization is well-designed to account for the ordinal characteristic of survival prediction, enforcing stable risk ordering across data shifts. Meanwhile, modality-decoupled adaptation is proposed to balance modal-wise adaptation across multi-modal data distributions. Additionally, An uncertainty-aware gradient correction is introduced to stabilize optimization. Extensive experiments demonstrate the effectiveness of SurvTTA in multi-modal survival analysis, enabling stable test-time adaptation and remarkable performance.

Links to Paper and Supplementary Materials

Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/1250_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{ZhaRon_SurvTTA_MICCAI2026,
        author = { Zhao, Rongchang AND Deng, Hengzhang AND Li, Shuo},
        title = { { SurvTTA: Order-Aware Test-Time Adaptation for Heterogeneous Domain Shifts in Multi-modal 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

    This paper proposes SurvTTA, a Test-Time Adaptation (TTA) framework designed specifically for multi-modal survival analysis in oncological prognosis. The authors address the critical challenge of heterogeneous domain shifts encountered when deploying models across different clinical sites. SurvTTA innovatively leverages the ordinal nature of survival risk by introducing a Topological Consistency module, a Decoupled Geometric Alignment module, and an Uncertainty-aware Gradient Correction module. Experiments on Head-Neck-PET-CT and HECKTOR datasets demonstrate that SurvTTA outperforms existing regression TTA methods by effectively maintaining risk-order consistency without requiring source data or target labels.

  • 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 work addresses the critical issue of domain shifts in multi-center medical deployments. By focusing on TTA without needing source data or target labels, it offers a practical solution for real-world clinical settings where data privacy and accessibility are major constraints.

    2.The core insight to leverage the ordinal nature of survival risk (rather than absolute values) is strong. The Topological Consistency module, which maintains both node-wise risks and edge-wise pairwise orderings, is a sophisticated method to preserve the survival structure under distribution shifts.

    3.The integration of Decoupled Geometric Alignment (handling modality-specific shifts) and Uncertainty-aware Gradient Correction (resolving optimization conflicts) demonstrates a comprehensive understanding of the instability issues in TTA, making the framework well-suited for complex heterogeneous data.

  • 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.Contradiction in Online Setting: Section 3.1 mentions an “online setting” with data discarded post-inference, yet the method runs “3 epochs” with rollback. Suggestion: Clarify how rollback works without stored data. Is a buffer used? This is crucial for reproducibility.

    2.The paper only compares against two specific regression TTA methods. Suggestion: Discuss or test against standard UDA/TTA classification methods (e. g. , adversarial learning, TENT) to better justify why existing general methods fail for survival analysis.

    3.The ablation study lacks analysis on key hyperparameters and augmentation strength. Suggestion: Add a brief sensitivity analysis, especially regarding 3D rotations, to ensure anatomical integrity isn’t compromised.

  • 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

    Equation (4), the derivation from the constrained optimization is dense. Suggestion: Briefly explain how the Lagrangian solution translates to the per-sample update and justify why prediction variance directly indicates objective conflict.

  • 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?

    It requires mandatory revision to fix the methodological description, clarify the online setting, and clean up the text. If these issues are not addressed, it could easily tip into rejection territory.

  • 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



Review #2

  • Please describe the contribution of the paper

    This paper proposes SurvTTA, a novel test-time adaptation (TTA) framework for multi-modal survival analysis under heterogeneous domain shifts .

    The key insight is that survival prediction is inherently ordinal, and that risk ordering remains invariant across domains, even under significant distribution shifts. Based on this observation, the authors introduce:

    • a topological consistency module to preserve risk ordering structure,
    • a modality-decoupled alignment strategy to handle heterogeneous shifts across modalities, and
    • an uncertainty-aware gradient correction mechanism to stabilize test-time optimization.

    The method enables adaptation without access to source data or target labels, and demonstrates strong improvements in cross-domain survival prediction across multiple datasets.

  • 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 highly relevant and challenging problem: cross-site generalization in multi-modal survival analysis under heterogeneous domain shifts. Unlike standard classification tasks, survival prediction has intrinsic ordinal structure, and the paper explicitly leverages this property. This is a well-grounded and insightful formulation. 2.A key strength is the idea of modeling survival prediction through risk ordering invariance, rather than absolute score matching. The proposed topological consistency formulation (node-wise + edge-wise constraints) provides a principled way to encode ordinal structure in test-time adaptation. This perspective is both novel and well aligned with survival analysis theory, and represents a meaningful contribution beyond existing TTA methods. 3.The experimental section is thorough and convincing. The use of risk ranking (rather than absolute prediction) and evaluation via Kaplan–Meier curves strengthens the clinical relevance of the work.

  • 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 compares with several TTA methods, the baselines are relatively limited. It would strengthen the paper to include comparisons with: more recent or stronger domain generalization / adaptation methods, or alternative survival-specific adaptation strategies. 2.The framework introduces multiple components (topology, alignment, gradient correction), which may increase computational overhead and optimization complexity. A discussion on efficiency and scalability would be helpful.

  • 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 idea of leveraging ordinal structure in survival analysis for TTA is particularly interesting and could be further explored in future work.
    • It would be valuable to investigate how the framework performs under more extreme domain shifts or with missing modalities.
    • A discussion on runtime cost and deployment considerations would further strengthen the practical relevance.
  • 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 presents a novel and well-motivated approach to test-time adaptation in multi-modal survival analysis, addressing a challenging and clinically important problem. The key strengths include a clear and insightful formulation based on survival ordering, strong methodological contributions, and convincing experimental validation. While there are minor limitations in baseline comparisons and analysis depth, these do not detract from the overall contribution. Overall, I find this work to be technically sound, novel, and impactful, and I would recommend acceptance.

  • 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 SurvTTA, a test-time adaptation (TTA) framework for multi-modal survival analysis under heterogeneous cross-site domain shifts. The core observation is that survival risk ordering is invariant across clinical sites and modalities, which is leveraged to adapt pretrained models using only unlabeled target data. The method is evaluated on two multi-site PET/CT datasets (Head-Neck-PET-CT and HECKTOR).

  • 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.Clinically well-motivated problem. TTA for survival analysis under heterogeneous multi-modal domain shifts is a genuinely underexplored problem with practical relevance for cross-site deployment. The paper is the first to address this combination. 2.The insight that survival risk ordering (rather than absolute scores) is site-invariant is well-motivated and directly informs the TC module design. Framing this as a topological graph with node- and edge-level consistency losses is technically coherent. 3.The DGA module’s per-modality treatment is a sensible and important design choice given that different imaging modalities (CT vs. PET) exhibit different degrees of domain shift. This is often overlooked in unified TTA approaches. 4.The failure of SSA on HMR and instability of ST variants highlight a real practical gap, and SurvTTA’s robustness across target sites is a meaningful result.

  • 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 HECKTOR target sets appear to have on the order of 20–30 patients each based on dataset sizes and splits. No confidence intervals or statistical significance tests are reported for any of the main results 2.The paper compares against only two regression TTA methods (CASP, SSA) and self-training variants. CASP is from sentiment analysis (AAAI 2025) and SSA from ICLR 2025 — neither is designed for survival prediction. Stronger baseline comparison is needed 3.ResNet-18 with cross-attention or Concat-MLP fusion is a fairly shallow source architecture. It is unclear whether performance gains reflect TTA effectiveness or compensating for a weak source model. 4.Table 2(b) shows that +T+G (mean 70.28%) underperforms +G alone (70.46%). This is a notable finding that contradicts the narrative and is dismissed in a single sentence. A thorough analysis of why these objectives conflict and why UGC resolves them is necessary to validate the design. 5.Presenting post-TTA significance (P=0.036) without confidence intervals or sample size . λ1, λ2, τ, c, and γ govern all three modules and their interactions, how is the hyperparameter selection done.

  • 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.

    (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 problem is well-motivated, and the core idea of leveraging rank-order invariance for survival TTA is interesting. However, the evaluation is critically underpowered, target sets of ~20–30 patients produce C-index estimates with very high variance, and no uncertainty quantification is provided for any result. The ablation reveals an internal inconsistency (TC+DGA < DGA alone) that is not adequately addressed. The baseline comparison is narrow and uses methods not designed for survival prediction. The UGC module has substantial overlap with CAGrad, but lacks sufficient differentiation. Significant additional experiments, baselines, and addressing queries are needed before the paper is ready for publication.

  • 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

N/A




Meta-Review

Meta-review #1

  • Your recommendation

    Provisional Accept

  • 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 test time adaptation (TTA) technique for survival analysis across different sites. The reviewers have acknowledged the importance of the problem and the soundness and practicality of the solution. They highlight three newly introduced components that were well described: a module to preserve the risk order, a modality alignment module, and an uncertainty-based stabilizer. Strengths include the relevance of the addressed problem (survival across sites), the evaluation done on two public datasets, and the improvement over a choice of baselines. Some issues were raised about the choice of baselines, as they rely on simple architectures and/or focus on natural image regression (instead of survival analysis). Another weakness is the lack of confidence intervals or a statistical significance test. Consider upgrading the paper to overcome these.



back to top