Abstract

Survival prediction from whole slide images (WSIs) is central to computational pathology, yet existing multiple instance learning (MIL) methods optimise exclusively for discrimination, with no explicit calibration objective. Miscalibration is especially harmful in clinical deployment, where systematic prediction errors across patient subgroups lead to inequitable risk estimates. We present MASurv, which augments discrete-time survival training with an Adversarial Conditional Moment Test (ACMT) on counting-process residual increments, directly encoding the calibration condition as a moment-orthogonality restriction decomposed locally in time within matched risk strata. A second ACMT task regularises within-patient prediction inconsistency for multi-slide patients. On our internal institutional cohort (662 patients, 5-fold cross-validation (CV)), MASurv achieves the best ICI = 0.096 and WG-ICI = 0.188, a 25% subgroup-calibration reduction over the NLL-only baseline (Cohen’s d = −1.54, very large effect). More critically, on zero-shot internal→TCGA-LUAD transfer (509 patients), MASurv is the only method retaining both discrimination (C-index = 0.680 ± 0.044) and calibration (D-cal p = 0.143, ICI = 0.079 ± 0.002); all baselines collapse (ICI > 0.36, D-cal p ≈ 0).

Links to Paper and Supplementary Materials

Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/4000_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{SabMeg_MASurv_MICCAI2026,
        author = { Sabouri Rad, Meghdad AND Huang, Vincent AND Hosseini, Mohammad Mehdi AND Nikbakhtsarvestani, Farzaneh AND Bixby, Amber AND Carello, Saverio J. AND El-Zammar, Ola AND Nasr, Michel R. AND Rodd, Bardia},
        title = { { MASurv: Risk-Conditional Martingale Adversarial Training for Calibrated Multi-slide Survival Prediction in Whole-Slide Histopathology Images } },
        booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
        year = {2026},
        publisher = {Springer Nature Switzerland},
        volume = {LNCS 16891},
        month = {September},
        page = {pending}
}


Reviews

Review #1

  • Please describe the contribution of the paper

    This paper proposes MASurv, a calibration-oriented survival prediction framework for whole-slide histopathology. The key idea is to enforce conditional calibration through adversarial conditional moment testing on counting-process residual increments, together with an auxiliary multi-slide consistency objective. Results on an internal SUNY cohort and zero-shot transfer to TCGA-LUAD suggest improved calibration, especially in terms of ICI, WG-ICI, and D-calibration, while maintaining competitive discrimination.

  • Please list the major strengths of the paper: you should highlight a novel formulation, an original way to use data, demonstration of clinical feasibility, a novel application, a particularly strong evaluation, or anything else that is a strong aspect of this work. Please provide details, for instance, if a method is novel, explain what aspect is novel and why this is interesting.

    The paper addresses an important problem in WSI survival modeling: calibration rather than pure ranking performance. The method is theoretically motivated by counting-process martingale residuals and evaluates both internal and external generalization. The inclusion of subgroup calibration analysis and ablations is appreciated.

  • 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 paper lacks comparisons with methods that explicitly optimize prediction calibration, such as SCMIL[1] and X-CAL[2] (the latter is also mentioned in the limitations section). 2.The writing and presentation of the paper would benefit from further improvement to enhance overall readability.

    [1] Yang Z, Liu H, Wang X. Scmil: Sparse context-aware multiple instance learning for predicting cancer survival probability distribution in whole slide images[C]//International Conference on Medical Image Computing and Computer-Assisted Intervention. Cham: Springer Nature Switzerland, 2024: 448-458. [2] Goldstein M, Han X, Puli A, et al. X-cal: Explicit calibration for survival analysis[J]. Advances in neural information processing systems, 2020, 33: 18296-18307.

  • 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?

    I lean weak reject because the problem is important and the external zero-shot calibration results are interesting. However, stronger baseline comparisons are needed to make and the paper writing requires optimization.

  • 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 #2

  • Please describe the contribution of the paper

    The main contribution is a calibration-oriented survival learning framework for WSI MIL based on counting-process theory. Specifically, the paper proposes: (1) A risk-conditional, time-local ACMT loss that enforces conditional moment orthogonality between residual increments and subgroup variables to improve calibration;

    (2) A multi-slide ACMT objective to regularize prediction consistency across slides from the same patient; (3) An empirical demonstration that the proposed training improves calibration and generalizes better under zero-shot cross-institution transfer compared to standard NLL-based MIL methods.

  • Please list the major strengths of the paper: you should highlight a novel formulation, an original way to use data, demonstration of clinical feasibility, a novel application, a particularly strong evaluation, or anything else that is a strong aspect of this work. Please provide details, for instance, if a method is novel, explain what aspect is novel and why this is interesting.

    (1) The problem is important and timely: moving beyond C-index to calibrated risk prediction, especially subgroup calibration, is clinically meaningful. (2) The method is well motivated: using martingale residuals for calibration is principled, and residual-level adversarial training is better justified than feature-level approaches. (3) The experimental protocol is strong: patient-level splits, 5-fold CV, zero-shot external evaluation, multiple calibration metrics, and informative ablations. (4) The external results are compelling: MASurv is the only method that maintains both discrimination and calibration under zero-shot transfer, which is highly relevant for deployment.

  • 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) Missing calibration-specific baselines: the lack of comparison with methods such as X-CAL or post-hoc recalibration makes it unclear whether gains come from ACMT specifically or from optimizing calibration in general. (2) Missing matched baseline on external data: the identical NLL-only baseline is not reported on TCGA, which weakens the claim that ACMT improves cross-institution calibration. (3) Limited generalization evidence: results are based on a single disease, dataset pair, and encoder, which is insufficient to fully support broad claims of cross-institution calibration transfer.

  • 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.

    (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 is a promising and potentially impactful paper addressing an important and underexplored problem in WSI survival modeling. The methodological idea is novel, and the external zero-shot results are encouraging. However, I cannot support a higher score in its current form due to several key limitations (as in major weaknesses part).

  • 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 experimental results are comprehensive.



Review #3

  • Please describe the contribution of the paper

    This paper proposes MASurv, a calibration-oriented survival MIL method for WSIs. The key idea is to use adversarial conditional moment testing (ACMT) on martingale residual increments to improve subgroup calibration and multi-slide consistency. The topic is important and the problem setting is interesting.

  • Please list the major strengths of the paper: you should highlight a novel formulation, an original way to use data, demonstration of clinical feasibility, a novel application, a particularly strong evaluation, or anything else that is a strong aspect of this work. Please provide details, for instance, if a method is novel, explain what aspect is novel and why this is interesting.

    The paper addresses a meaningful gap. Most prior WSI survival work optimizes NLL and reports C-index, whereas this paper explicitly targets calibration and worst-group calibration. That is a worthwhile direction and clinically relevant. The external zero-shot TCGA results further underscore the importance, as MASurv achieves much better ICI/WG-ICI and is the only method with non-rejected D-calibration in the external 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.

    The paper omits important details, which hurts clarity and makes the method harder to follow. The related work section feels incomplete and lacks a clear logical flow, especially in motivating the adversarial design. The experiments miss the relevant calibration baseline, which weakens the empirical support for the main claims.

  • 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?

    The manuscript explicitly names the source institution for the internal cohort. In the paper, it shows as “SUNY (internal cohort). 662 lung cancer patients, SUNY Upstate Medical University” on Page 5.If this dataset is not publicly available (I’m not sure if it is a public dataset), could this potentially raise a double-blind anonymization concern under MICCAI policy? This may be worth checking.

  • 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

    Here are some questions: 1.Eq. (1) is introduced without notation explanation, which hurts readability; moreover, the claim that systematic non-zero correlation is “equivalent” to subgroup miscalibration appears too strong and would need a more careful justification. 2.The motivation for the adversarial formulation is insufficiently explained. The paper states that calibration is enforced adversarially, but does not clarify why an adversarial critic is necessary or preferable to simpler moment-matching penalties. 3.The related-work section is not well connected, especially the adversarial-training paragraph, which distinguishes the method from domain-adversarial learning but does not clearly motivate the proposed adversarial design. 4.Fig. 1 is hard to parse and appears inconsistent with the main text (“stage” in the figure vs. “slide count” in Sec. 3.3). 5.The method uses sex, ECOG, and slide count as nuisances, and the evaluation also emphasizes groups based on sex / ECOG / slide-count. But the data section says the cohort includes other covariates such as age and Charlson Comorbidity Index. Why were these three chosen, and why were others excluded? Or did I miss something? 6.The optimization objective is not written rigorously: Eq. (7) suggests a single minimization, while Sec. 3.3 states that the critic maximizes the ACMT objective. The roles of the last two terms are not clearly explained. In particular, the paper does not make it sufficiently clear how L_consist and L_ent are defined, why they are needed, and how they interact with the main ACMT objective. 7.X-CAL is discussed as a relevant prior calibration method, yet the empirical comparison is deferred without sufficient justification.

  • 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?

    My score is mainly driven by three concerns: an unclear motivation for the adversarial training design, limited clarity and rigor in the method presentation, and the lack of direct comparison with other relevant calibration methods. Although the problem is important and the results are promising, these issues make the current version less convincing than it could be.

  • 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.

    I lean toward borderline accept. The rebuttal addresses most of my concerns, especially regarding notation, adversarial motivation, covariate choice, and the optimization objective. However, my support depends on whether the camera-ready version actually incorporates the promised clarifications, as several technical details remain insufficiently clear in the current submission.



Author Feedback

We thank the Reviewers and Area Chair. We address AC’s two priorities first. 1.(AC-1, R1-W1, R2-W1, R4-Q7) X-CAL, SCMIL, post-hoc recalibration. Post-hoc recalibration is not designed for subgroup-conditional calibration under censoring and typically requires held-out calibration data from the deployment distribution, unavailable in zero-shot transfer (§4.3). X-CAL [15]: Goldstein et al. ([15] §6) state conditional calibration is “in general infeasible even to measure (let alone optimize) … without strong assumptions”; X-CAL is a marginal D-calibration proxy with no time-local moment and is not formulated for multi-slide MIL; adapting it would test adaptation choices, not the calibration principle. ACMT targets a different statistical object: conditional residual-moment orthogonality at risk-stratum level. SCMIL [Yang MICCAI’24] trains censored NLL on a Gaussian-mixture head (Eq. 9) with no calibration objective in its training loss; its contributions are architectural (SoftFilter, SCSA sparse attention, RegisterMDN), not a residual-calibration objective comparable to ACMT. 2.(AC-2, R4-Q2) Adversarial motivation. Conditional calibration requires E[ΔM̂|f(W)]=0 for every measurable f, an infinite family of moment constraints (GMM/conditional-moment view, §3.3, [8]). A scalar fixed-direction penalty enforces only one such f and misses non-linear and interaction violations; the critic D_ψ implements a worst-case test over its parametric class. Fig. 5 (§4.4) corroborates: ACMT→BCE raises ICI; Feature-adversary on z_i (closest DAT analogue) gives the worst WG-ICI; Global-moment and No-conditioning degrade ICI; No-adversary raises ICI. The form of the objective, not its mere presence, drives the gain. 3.(R2-W2) Matched NLL-only baseline on TCGA. Table 1 isolates mechanism on SUNY: NLL Baseline ICI=0.117, WG-ICI=0.249 vs. MASurv 0.096/0.188.On TCGA, three NLL-trained MIL baselines span architecture at constant loss (ABMIL ICI=0.392, DTFD-MIL 0.366, DeepAttnMISL 0.382; D-cal p≈0); MASurv (ICI=0.079, D-cal p=0.143) remains calibrated. With SUNY↔TCGA Spearman ρ=−0.70, the gap supports lack-of-calibration-objective over architecture-specific failure as the explanation. We agree the exact λ_ACMT=0 row on TCGA would be informative. 4.(R2-W3) Generalization scope. SUNY (n=662) and TCGA-LUAD (n=509) span institution, scanner, population, and label process: the principal calibration-relevant shift axes. We acknowledge a single disease (LUAD) and single encoder; cross-encoder (UNI, GigaPath) and cross-cancer extensions are future directions (§5). 5.(R4-Ethics) Anonymization. The cohort is de-identified under institutional protocols; the institution name was data provenance only and will be relabeled as an internal institutional cohort in camera-ready. No author names, grant IDs, or identifying acknowledgments are present. 6.(R4-Q1, Q3, Q4, Q5, Q6 and R1-W2) Method clarity and presentation. L_consist enforces stop-gradient consistency between patient hazards λ̂i(k) and slide-level hazards λ̂{i,s}(k); L_ent is an attention-entropy regularizer on w_{i,s}. The min-max in §3.3 applies to the ACMT terms (Eqs. 5–6) within the summed loss (Eq. 7): critics D_ψA, D_ψB maximise; the model minimises Eq. 7.Sex, ECOG, and slide count cover demographic, clinical, and multi-slide axes; age and CCI overlap with legitimate hazard signal and were excluded to avoid orthogonalizing against true risk. DAT acts on representation features as a feature-invariance property; ACMT acts on residual increments ΔM̂_i as a calibration property — different layers, different statistical objects, not interchangeable; Fig. 5’s Feature-adversary is the DAT analogue and gives the worst WG-ICI. “Stage” in Fig. 1 is a typo for “slide count” (§3.3); we will correct this. Eq. 1: ΔM̂_i(k)=ΔN_i(k)−Y_i(k)λ̂_i(k) (ΔN bin event indicator, Y at-risk, λ̂ per-bin hazard); “equivalent” in §1 will be softened to “implies”.




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 an approach for multi-slide survival prediction for histopathology image analysis with improved calibration.

    Reviewers highlight the importance of model calibration and the overall motivation within the paper. Ablation studies are seen as valuable, and external results for zero-shot transfer as seen as strong.

    Reviewers are missing comparisons with explicit prediction calibration methods such as X-CAL or SCMIL, including for zero-shot prediction on the external cohort. Additionally, the motivation of the adversarial method is not fully clear.

    In their rebuttal, authors should more clearly explain why no comparison was conducted with methods such as X-CAL or with post-hoc recalibration. They should further explain the motivation for the adversarial formulation more clearly. Of note, the goal of the rebuttal should not be to add additional comparison methods or similar, as these results cannot be sufficiently assessed by the reviewers within the rebuttal process.

  • 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 rebuttal addressed important questions regarding methodological details and comparison methods, with a reasonable reply also to reviewer #1 who did not update their assessment. I therefore lean towards accept.



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.

    Reasonable reviewer comments have been addressed.



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.

    Thank you the authors for the detailed rebuttal and clarifications. The rebuttal addressed many concerns effectively, particularly around the motivation for adversarial calibration, positioning relative to X-CAL and SCMIL, methodological clarity, and the distinction between feature invariance and residual calibration. The experimental evaluation was generally viewed as comprehensive, and the proposed objective is supported by consistent ablation evidence.

    Reviewers acknowledged improved clarity and stronger positioning after rebuttal. Remaining concerns are largely about presentation and completeness of technical explanations rather than fundamental validity of the approach, and I recommend the authors to address these appear in the camera-ready version.

    Overall, the paper presents a technically sound and well-motivated contribution with convincing empirical evidence for robust calibration under cross-institutional transfer. I recommend Accept.



back to top