Paper Info Reviews Author Feedback Meta-Review Back to top
List of Papers Browse by Subject Areas Author List

Abstract

Whole-slide image (WSI) analysis is foundational to digital pathology, yet granular four-class metastasis grading (Normal, ITC, Micro, and Macro-metastasis) remains inherently challenging. Existing pseudo-bag strategies typically aggregate features from all generated bags for classification. However, such stochastic grouping inevitably introduces severe semantic dilution and label noise, especially in sparse categories like ITC and Micro-metastasis, where numerous pseudo-bags may be entirely void of discriminative signals. To mitigate this, we propose a novel tiered MIL framework designed to strategically filter out class-irrelevant noise. Specifically, a Progressive Shapley-based module is introduced to quantify the diagnostic contribution of each patch, facilitating the construction of ordered, high-quality pseudo-bags. Subsequently, a Reward-Guided Selection Module (RGSM) is employed to estimate pseudo-bag labels and dynamically discard those exhibiting conflicts with the global slide-level ground truth. Extensive experiments on a newly curated four-class WSI dataset demonstrate that our purification strategy achieves state-of-the-art performance. Notably, it significantly enhances the detection of highly sparse metastasis, outperforming the best baseline by 10.66% in F1-score.

Links to Paper and Supplementary Materials

Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/1020_paper.pdf

SharedIt Link: Not yet available

SpringerLink (DOI): Not yet available

Supplementary Material: Not Submitted

Link to the Code Repository

https://github.com/PS-zhang/RGSM

Link to the Dataset(s)

NIMM dataset: https://pan.baidu.com/s/1b8Tww2drwT5b7MZpBdfyGA password: nsrb

BibTex

@InProceedings{JiaQi_RewardGuided_MICCAI2026,
        author = { Jia, Qi AND Zhang, Pengshuo AND Yao, Shuilian AND Liu, Yu AND Sun, Lili AND Zhu, Yanmei AND Fan, Xin},
        title = { { Reward-Guided Distillation: A Progressive Pseudo-Bag Purification Framework for WSI Multiple Instance Learning } },
        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 a reward-guided distillation framework for WSI multiple instance learning, focusing on pseudo-bag purification to reduce noise and semantic dilution. It introduces a Progressive Shapley-based construction and a Reward-Guided Selection Module to filter unreliable pseudo-bags. Experiments show improved performance, especially for sparse metastasis classes.

  • 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 critical issue in WSI MIL, namely semantic dilution caused by noisy pseudo-bags, which is especially important for sparse categories like ITC and Micro-metastasis. 2.The Reward-Guided Selection Module introduces a novel generator-rewarder paradigm to evaluate pseudo-bag quality and filter noisy samples dynamically. 3.The paper demonstrates clear improvements in F1-score and AUC.

  • 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 Shapley-based ranking mechanism may introduce significant computational overhead, especially for large-scale WSI datasets with thousands of patches. 2.The paper does not analyze the sensitivity of key hyperparameters, such as the number of pseudo-bags and reward thresholds. 3.How does the proposed strategy compare with recent large-scale WSI foundation models?

  • Please rate the clarity and organization of this paper

    Good

  • Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.

    The submission does not mention open access to source code or data, but provides a clear and detailed description of the algorithm to ensure reproducibility.

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

    N/A

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

    N/A

  • Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.

    (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?

    The paper addresses an important problem in WSI multiple instance learning by reducing noise and semantic dilution in pseudo-bags. The proposed Shapley-based construction and reward-guided selection are well-motivated and show clear empirical improvements, especially for sparse classes. However, framework is relatively complex with limited validation on computational complexity.

  • 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 addressed my concerns, I recommend accept.



Review #2

  • Please describe the contribution of the paper

    The main contribution is a novel tiered MIL framework with progressive pseudo-bag purification, named Reward-Guided Distillation. It consists of two core components: a Progressive Shapley-based module to rank patches by diagnostic contribution and construct ordered pseudo-bags, and a Reward-Guided Selection Module (RGSM) with a generator-rewarder structure to score and filter low-quality pseudo-bags via cosine-similarity-based reward signals. The method achieves state-of-the-art performance on the newly curated NIMM dataset and significantly improves the detection of sparse metastatic lesions.

  • 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 practical, addressing the real-world bottleneck of noisy pseudo-bags in multi-class WSI classification, which is valuable for digital pathology applications.
    • The proposed framework is well-designed, combining Shapley-value-based instance ranking and reward-guided bag filtering, showing a systematic solution for pseudo-bag purification.
    • The RGSM module is technically sound, using cosine similarity to define reliable reward signals and providing a learnable way to remove class-irrelevant bags.
  • 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 definition of reward is strongly supervised by slide-level ground truth through pseudo-bag relabeling, which limits its flexibility and generalization to fully unsupervised or more weakly supervised scenarios.
    • The experimental evaluation is only conducted on the self-collected NIMM dataset. Performance on public WSI benchmarks (e.g., CAMELYON16/17) is missing, so the generalizability of the method is not fully verified.
    • Some implementation details are insufficiently described, such as the exact procedure of Shapley value approximation in large-scale WSI patches and the hyperparameter selection strategy for α and λ.
  • 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 paper addresses an important clinical task with a clear and empirically effective framework. The proposed pipeline improves pseudo-bag quality and boosts performance on sparse metastasis categories. However, the methodological novelty is moderate, the generalization is not validated on public datasets, and some technical descriptions and novelty claims need refinement.

  • Reviewer confidence

    Somewhat confident (2)

  • [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 partially addressed my concerns. Therefore I will keep my score as weak accept.



Review #3

  • Please describe the contribution of the paper

    This paper proposes a novel tiered MIL framework designed to strategically filter out class-irrelevant noise. A Progressive Shapley-based module is introduced to quantify the diagnostic contribution of each patch, facilitating the construction of ordered, high-quality pseudo-bags. Subsequently, a Reward-Guided Selection Module (RGSM) is employed to estimate pseudo-bag labels and dynamically discard those exhibiting conflicts with the global slide-level ground truth.

  • 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.

    Reasonable Architectural Design: The proposed framework first employs an instance ranking and reorganization module to construct systematically ordered pseudo-bags based on their diagnostic contribution. Subsequently, they propose a Reward-Guided Selection Module (RGSM) consisting of a pre-trained Generator and a Rewarder. The Generator estimates pseudo-labels for each bag, while the Rewarder calculates a reward score based on the cosine similarity between pseudo-labels and ground truth slide labels. This smooth, well-calibrated metric allows the framework to dynamically discard class-irrelevant bags, effectively eliminating redundant noise before final classification.

  • 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.Insufficient details regarding grammar and word choice. Regarding the title, “Reward-Guided Distillation: A Progressive Pseudo-bag Purification Framework for WSI Multiple Instance Learning”, is “Reward-Guided Distillation” intended to summarize or represent “A Progressive Pseudo-bag Purification Framework”? 2.The grammar throughout the manuscript requires further refinement. For example, on Page 1: “As WSI classification is inherently a weakly supervised problem—where a single slide-level label corresponds to a vast bag of instances—the core challenge lies in effectively identifying and aggregating discriminative patches that truly represent the underlying pathology [11].” This sentence should be carefully revised for clarity, conciseness, and grammatical accuracy. 3.Inconsistent terminology is used for the proposed modules and framework. For example, on Page 2, the phrase “instance ranking and reorganization module” in “Specifically, our framework first employs an instance ranking and reorganization module to construct systematically ordered pseudo-bags based on their diagnostic contribution” is inconsistent with the “Progressive Shapley-based module” mentioned in the abstract. Similarly, the framework is referred to as “the first class-agnostic pseudo-bag screening framework” in the contribution statement, whereas the title uses “Pseudo-bag Purification Framework.” The terminology should be unified throughout the manuscript, and the name used in the title should be consistent with that used in the contributions. 4.The methodological novelty appears to be limited. The proposed framework mainly integrates existing components, namely an instance ranking and reorganization module and a reward-guided selection module, for feature purification. The method appears to rely on pre-extracted features and performs screening rather than deeper feature mining. The main innovation I can identify is the use of a cosine-similarity-based reward score to isolate high-fidelity signals and prune non-contributory pseudo-bags. The authors should further clarify and strengthen the methodological novelty of the proposed approach.。 5.Figure 1 lacks subfigure labels. There are no labels such as “(a),” “(b),” and “(c)” in Figure 1.Please add them where appropriate and ensure that they are consistent with the figure caption and in-text references. 6.The module naming is inconsistent. For example, on Page 3, the manuscript states: “These features are then evaluated via a Progressive Shapley-based module to quantify the diagnostic contribution of each patch, facilitating the construction of ordered pseudo-bags.” Please verify whether the term should be “Progressive Shapley-based module” and ensure that it is consistent with the terminology used in the abstract and throughout the manuscript. 7.Please verify the correctness of Equation (1). In particular, the prediction function f(\cdot) is not clearly defined. The authors should provide a more explicit explanation of its input, output, and role in the formulation. 8.Please check whether {\rm \operatorname{MLP}}_G(f;\phi)on Page 4 is incorrect. Should it be {\rm \operatorname{MLP}}_G(b;\phi) instead? 9.The manuscript frequently switches among the terms “feature,” “embedding,” and “representation.” Please carefully check and standardize the usage of these terms throughout the manuscript to avoid ambiguity. 10.There is a grammatical issue in the Implementation Details section on Page 6.For example: “Experiments are conducted on two RTX 3060 GPUs for 200 epochs using Adam optimizer with lr = 10−4, 10−4 weight decay and a 20-epoch early stopping patience.” This sentence should be revised. A possible version is: “Experiments were conducted on two RTX 3060 GPUs for 200 epochs using the Adam optimizer with a learning rate of {10}^{-4}, a weight decay of {10}^{-4}, and an early-stopping patience of 20 epochs.” 11.The class distribution and Table 1 require further explanation. The manuscript states: “The final four-class annotation distribution consists of 231 Negative cases, 40 ITCs, 76 Micro-metastasis cases, and 282 Macro-metastasis cases.” However, Table 1 only includes ITC and Micro. Please explain this discrepancy. In addition, the expression “Mean ± Standard Deviation” is inconsistent with the way the results are presented in the table and should be revised accordingly. 12.Spacing errors should be carefully checked throughout the manuscript. For example, on Page 7: “By evaluating the cellular status within each patch, HoVer-Net can reliably derive the patch-level ground truth labels.As illustrated in Fig. 3(B), where the proportion is calculated as the average ratio….” There is a missing space between “labels.” and “As.” Similar formatting and spacing issues should be systematically corrected 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 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.

    (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 proposes a novel tiered MIL framework designed to strategically filter out class-irrelevant noise. However, the methodological novelty is limited.

  • Reviewer confidence

    Confident but not absolutely certain (3)

  • [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.

    Reject

  • [Post rebuttal] Please justify your final decision from above.

    The methodological novelty appears to be limited. The main innovation I can identify is the use of a cosine-similarity-based reward score to isolate high-fidelity signals and prune non-contributory pseudo-bags. The authors should further clarify and strengthen the methodological novelty of the proposed approach.



Author Feedback

We sincerely appreciate the reviewers’ insightful comments. We are also grateful for the recognition of our performance gains, research motivation and methodological design, whose novelty is described as “a novel tiered MIL framework designed to strategically filter out class-irrelevant noise.”

[Shapley-based approach (R2,3,4)] The core of our Shapley-based approach lies in quantifying the marginal contribution of each patch to the overall category prediction. (1) Marginal Contribution Calculation: The exact Shapley value evaluates a patch’s impact by inserting it into all possible subset combinations. In Eq. (1), the function f(⋅) maps subset features to a WSI-level predictive probability. A patch’s marginal contribution is the predictive difference before and after adding it to a subset. Larger differences indicate greater diagnostic importance, yielding higher Shapley scores. (2) Monte Carlo Approximation: To avoid intractable exponential complexity, we replace exhaustive traversal with the stochastic sampling of M background subsets. This effectively reduces the computational overhead to a linear scale. (3) Estimation Stability: To ensure approximation precision, we perform T independent sampling iterations. The final Shapley score is averaged across M times and T trials, providing a stable ranking of patch importance for pseudo-bag construction.

[Hyperparameter Settings (R2,3)] (1) Pseudo-bag Count: As indicated in Fig. 3(A), the optimal pseudo-bag count is set to 10.(2) Reward Threshold: The threshold was optimized over [0.2, 0.8] with a 0.05 step size. AUC ranged from 0.9154 to 0.9584, achieving its peak performance at 0.5.(3) Objectives (α and λ): Both hyperparameters were optimized over [0, 1] with a 0.1 step size. Overall AUC fluctuated between 0.8997 and 0.9584, peaking at α=0.5 and λ=0.4. [Compare with large-scale WSI foundation model (R2)] Our method demonstrates superiority over general foundation models. We evaluated the foundation model GigaPath (Nature, 2024) on our NIMM dataset. GigaPath achieved an overall AUC of 95.63% and F1-score of 71.12%, underperforming our method by 0.98% and 9.29%. For sparse categories, GigaPath yielded F1-scores of 43.48% (ITC) and 48.28% (Micro), underperforming our method by 25.45% and 29.02%.

[Reward and Generalizability (R3)] WSI classification is a weakly supervised problem lacking patch-level labels. Therefore, our reward, guided by slide-level labels, operates in this weakly supervised paradigm. Unsupervised extensions are reserved in future. Experiments on the CAMELYON17 demonstrate our framework’s generalizability. Our method yielded F1-scores of 58.52% (Overall), 43.14% (ITC) and 52.94% (Micro), outperforming the second-best baseline MRePath (57.68%/42.86%/49.52%) by 1.46%, 0.65% and 6.91%.

[Novelty and Other Issues (R4)] “Reward-Guided Distillation” encapsulates our core contribution: achieving feature distillation through the first class-agnostic pseudo-bag screening framework. By strategically filtering out irrelevant bag-level features, we directly address the feature drift caused by noisy instances in multi-class WSI analysis. (1) We introduce a Shapley-based module (accelerated via Monte Carlo approximation) to establish strong patch-category correlations. (2) We propose a novel Generator-Rewarder architecture to dynamically quantify pseudo-bag quality during training. (3) We establish a closed-loop mechanism ensuring that highly sparse, latent pathological signals (e.g., ITC and Micro) are deeply isolated and amplified for classification, advancing beyond standard feature screening. We highlighted ITC and Micro due to the extreme sparsity and diagnostic challenge. Our method achieved top-tier overall performance, yielding AUC/F1-scores of 99.97%/98.95% (Negative) and 98.88%/94.12% (Macro), outperforming the second-best method (MRePath) by 0.1%/0.19% and 1.25%/1.30%, respectively. We will conduct a thorough overhaul to rectify formatting, among others.



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 reward‑guided distillation framework for WSI multiple instance learning that uses a Progressive Shapley‑based module to rank patches and a Reward‑Guided Selection Module to filter noisy pseudo‑bags, improving classification especially for sparse metastasis classes. The reviewers agree that the framework is well-motivated and the architectual design is reasonable. The reviewer also recognized the empirical improvements. Concerns are raised on the missing public benchmarks, insufficient implementation details and limited methodological novelty. Please respond to the main concerns of the reviewers, including the following ones:Clarify the exact procedure of Shapley value approximation,clarify the methodological novelty,why does the text state a four‑class distribution (Negative, ITC, Micro, Macro), but Table 1 only include ITC and Micro.

  • 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 paper received two accept and one reject after rebuttal. I tend to reject the paper because of the following two reasons. First, the method is only evaluated on one small newly constructed dataset and its generalibility is unknown. Second, pathology foundation medels have been widely used in WSI analysis for patch feature extraction, which generally results higher performance than training Resenet50.But this paper is based on ResNet50 for patch feature extraction.



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 reviewers generally agree that the paper addresses an important problem with a well-motivated framework and demonstrates solid empirical improvements. After rebuttal, key technical concerns were mostly addressed, resolving most major uncertainties.



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.

    After rebuttal, this paper received two accept recommendations and one rejection. The AC acknowledges that the paper addresses an important and clinically relevant problem in WSI multiple-instance learning: reducing noisy pseudo-bags and mitigating semantic dilution for sparse metastasis categories. The proposed reward-guided distillation framework, which combines Progressive Shapley-based patch ranking with a Reward-Guided Selection Module, is well motivated and reasonably designed. Reviewers also recognize its empirical improvements in AUC and F1-score, particularly for the challenging ITC and micro-metastasis classes. The AC believes that the authors have addressed most of the concerns raised during review. Regarding methodological novelty, the AC considers the work sufficiently novel, though the authors are encouraged to strengthen the presentation in the final version. Therefore, the final recommendation is accept.



back to top