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

Abstract

Continual learning (CL) for pathological whole slide image (WSI) classification is challenging because gigapixel images, sparse diagnostic cues, and cross-organ domain shifts exacerbate catastrophic forgetting. In addition, existing methods often rely on rehearsal buffers, which are difficult to deploy in clinical practice under privacy and governance constraints. We propose knowledge memory pool multiple instance learning (KMP-MIL), a rehearsal-free continual MIL framework that couples a frozen pathology foundation model (FM) with parameterized memory-based adaptation. KMP-MIL introduces: 1) a growing KMP that incrementally stores compact task-specific memory units instead of replayed samples, and 2) query-based memory retrieval with prototype-conditioned feature calibration (PFC) and textual calibration (TC) in a shared vision-language space to improve adaptation. The framework supports task-CL with task identity (ID) and class-CL with task-agnostic inference through a universal memory pool that retrieves task-relevant knowledge without task ID. Experiments on sequential cross-organ WSI classification across breast, lung, kidney, and esophagus datasets show strong and stable performance. KMP-MIL achieves 89.08% average accuracy in task-CL and competitive class-CL results with reduced forgetting. It is also stable under reverse-order evaluation and storage-efficient, with persistent memory growing from 0.80 MB to 3.21 MB over four tasks, about 3.14% of a replay buffer storing 10 WSIs. Our code is available at https://github.com/Lingling-Yuan/KMP-MIL.

Links to Paper and Supplementary Materials

Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/2180_paper.pdf

SharedIt Link: Not yet available

SpringerLink (DOI): Not yet available

Supplementary Material: Not Submitted

Link to the Code Repository

https://github.com/Lingling-Yuan/KMP-MIL

Link to the Dataset(s)

BRACS dataset: https://www.bracs.icar.cnr.it/ TCGA datasets: https://portal.gdc.cancer.gov

BibTex

@InProceedings{YuaLin_KMPMIL_MICCAI2026,
        author = { Yuan, Lingling AND Yin, Zhaoxia AND Han, Yan AND Zhang, Jinghua AND Grzegorzek, Marcin AND Li, Chen},
        title = { { KMP-MIL: Knowledge Memory Pool Multiple Instance Learning with Foundation Model for Continual Whole Slide Image Classification } },
        booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
        year = {2026},
        publisher = {Springer Nature Switzerland},
        volume = {LNCS 16878},
        month = {September},
        page = {pending}
}

Reviews

Review #1

  • Please describe the contribution of the paper

    This paper addresses the critical issues of catastrophic forgetting and the inability to use sample replay under privacy constraints in the context of continuous classification of whole-slide pathology images (WSI). It proposes a replay-free knowledge memory pool multi-instance learning framework, KMP-MIL. By integrating a frozen pathological base model, an incremental Knowledge Memory Pool (KMP), Prototype-Conditional Feature Calibration (PFC) and Text Calibration (TC), this method enables cross-organ continuous learning and supports both known and unknown inference modes for the task.

  • 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.

    By combining a knowledge memory pool, feature calibration, vision-language alignment and weakly supervised MIL, the training process is logically sound and effectively mitigates cross-domain drift and catastrophic forgetting.

    It supports both Task-CL (known task IDs) and Class-CL (unknown task IDs), bringing it closer to real-world clinical deployment scenarios.

  • 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. Each task is restricted to a fixed number of M=10 memory units; it is not possible to adaptively increase or decrease this number based on task complexity or the number of categories, nor are there any mechanisms for memory merging, deduplication, compression or pruning. As the number of tasks continues to increase, the memory pool may expand linearly, thereby affecting the model’s scalability.
    2. Figure 1: The block diagram is visually appealing but lacks sufficient annotation of key processes; it is recommended to supplement it with explanations regarding data flow and the input/output dimensions of modules.
    3. The data flow between modules is not seamless; it is recommended that relevant details be added to ensure greater clarity for the reader.
    4. Notation is inconsistent, with the same variable written differently in different sections (e. g. , ‘M’ in the Methods section and ‘m’ in the Experiments section). The authors are advised to carefully review and correct this.
    5. There are a large number of hyperparameters (M/K/μ/τ/λ_route/λ_reg/λ_sim), yet insufficient sensitivity analysis and justification for parameter selection have been provided.
    6. The analysis of the experimental results lacks depth. Section 3.2 ‘Results’ primarily presents data without extensive analysis; in the ablation experiments, only performance curves are provided, with no analysis of the underlying mechanisms or reasons for the results.
    7. In Table 2, the evaluation metrics for the ‘Task-CL setting’ and ‘Class-CL setting’ tasks should be consistent.
  • 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?

    The research demonstrates a degree of innovation, the methodology is clearly described, and the experimental work is thorough.

  • 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

    A framework for whole slide image (WSI) multiple instance learning (MIL) that eliminates need for data replay in continual learning by story task-specific memory units. It uses a query-based memory retrieval scheme that combines PFC and TC. It is evaluated on continual learning sequentially across 4 datasets and demonstrates good results robust to reverse ordering.

  • 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.
    • leverages VLM foundation model to unify text prompt with bag representation tokens for more accurate retrieval
    • Two calibration schemes that combines the structure of tailored information storage with flexible for memory unit enhancement.
    • Effective solution for memory efficiency and data privacy concerns not allowing storage of past training data/per sample level features
  • 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.
    • For interpretability’s sake, it would be interesting to explore for class-CL, how many of the Top-K come from the correct task.
    • It would be beneficial to justify the choice of max pooling as the method to obtain the query and whether this compression is too strong as it might loss structural information in the slide. This is especially important as the entire pipeline relies on queried retrieval which is dependent upon cosine similarity between the query and the memory token.
    • While the reverse order ablation show the robustness of this method, I wonder if it is sufficient to show the effectiveness of this model over more datasets with even more different anatomical structures. This is especially interesting as the “no-replay” design after training each task ensures that earlier tasks knowledge store cannot be refined, so order directly impacts the quality of knowledge memory pool.
  • 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?

    While the quantitative results are strong, it would be appreciated to have more quantitative justification for the design of the framework. Moreover, given the limited nature of the tested dataset, more arguments on the generalisability of this method is appreciated.

  • 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 #3

  • Please describe the contribution of the paper

    This paper addresses catastrophic forgetting in continual learning for pathological whole slide image classification and proposes a replay-free framework, KMP-MIL. The method builds a progressively growing memory pool and combines memory retrieval, feature calibration, and textual calibration to achieve solid performance and storage efficiency in cross-organ sequential learning.

  • 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 problem setting is meaningful and practically relevant, especially considering that replay is often restricted in medical scenarios. The overall method is also fairly well designed, as it does not simply add a memory module but integrates retrieval, calibration, and vision-language prototypes into a unified framework. In addition, the experiments are relatively comprehensive, covering task-CL, class-CL, reverse-order evaluation, and storage cost, which gives a reasonably complete picture of the method’s effectiveness.

  • 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.

    There are still several notable weaknesses. First, the improvement over strong baselines is relatively modest overall. While the results are positive, the margin is not yet fully convincing, and it would be helpful to include additional stability experiments to show that the gain is robust. Second, the paper would benefit from more intuitive visualizations, such as which memory units are retrieved, or how the features/prototypes change before and after calibration. Such analyses would make the method easier to understand. Finally, the explanation of the memory retrieval and calibration mechanisms is still somewhat limited. At present, the paper mainly shows that these components help empirically, but the analysis of why they help and what role they actually play could be made deeper and clearer.

  • 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?

    Insufficient analysis of experimental results

  • 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



Author Feedback

We thank the meta-reviewer (M) and reviewers (R1–R3) for their constructive feedback. We are encouraged by their positive comments on our method and manuscript clarity, and respond to the main concerns below. 1.Scalability and task-order robustness (R1-1, R2-3, R3, M). KMP uses a fixed budget of M=10 memory units per task, not adapted to task complexity or class number. Thus, the pool grows linearly with tasks. In the final version, we will clarify this limitation and discuss adaptive allocation, merging, deduplication, pruning, and compression as future directions. KMP stores compact task-level units rather than historical WSIs, patch features, or replay samples. In our four-task benchmark, storage grows from 0.80 MB to 3.21 MB, 3.14% of a replay buffer with 10 WSIs. Since earlier memories are frozen, task order may affect memory quality. We treat reverse-order evaluation as an order-sensitivity check, not proof over all permutations. With 10-fold cross-validation, it gives initial evidence of stability on four organs. Longer streams, broader anatomy, and more orders remain future work. The modest gains are meaningful under strict no replay while reducing forgetting and storage cost. 2.Hyperparameters and ablation (R1-5, R1-6, R3, M). Fig. 2D–G focuses on four key hyperparameters: M, K, λ_route, and λ_reg. M and K balance memory cost and retrieval coverage: small values may miss useful knowledge, while large values introduce redundant memories. λ_route improves task-agnostic routing, and λ_reg limits over-modulation in PFC. μ and τ define the margin and temperature of L_route, while λ_sim encourages semantic consistency. In the final version, we will strengthen the hyperparameter justification and ablation discussion. Specifically, KMP preserves task-specific knowledge, L_route reduces routing ambiguity in class-CL, PFC and L_reg calibrate cross-organ feature shifts, and TC adds semantic alignment. One default setting was used across all experiments, and the sensitivity curves are stable around the selected settings. 3.Query design and retrieval interpretability (R2-1, R2-2, R3, M). In class-CL, task ID is not used; memory units are retrieved from the universal KMP by query-key cosine similarity. Checking whether Top-K units come from the correct task is a useful post-hoc diagnostic, not part of inference. In the final version, we will discuss Top-K task-origin tracing as a future interpretability analysis. For max pooling, q_i is only a lightweight retrieval router, not the final slide representation. Prediction still uses calibrated patch-level features, prototype-guided MIL aggregation, and TC-enhanced class prototypes, so instance-level information is modeled downstream. Max pooling is chosen because WSI diagnosis often depends on sparse high-response regions and avoids extra trainable query modules. 4.Mechanistic interpretation and visualization (R1-6, R3, M). KMP provides knowledge-conditioned context without replaying historical WSIs. L_route reduces routing ambiguity in class-CL. PFC calibrates cross-organ feature shifts, while L_reg prevents overly strong modulation. TC adds complementary semantic alignment. In the final version, we will strengthen the Results and Ablation discussion to clarify each component beyond performance changes. Visualizing retrieved memories and feature/prototype shifts before and after PFC/TC would further improve interpretability, and we will highlight these analyses as future extensions. 5.Presentation, notation, and metrics (R1-2, R1-3, R1-4, R1-7, M). In the final version, we will revise Fig. 1, its caption, the Method description, and notation to clarify data flow and keep M consistent. We will also clarify the Evaluation Metrics section: ACC, AUC, and PR-AUC are classification metrics, and the corresponding forgetting rates are reported when the base metrics are included, especially for class-CL where task ID is unavailable and cross-task confusion is central.



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 is within scope for MICCAI and addresses a practically important pathology problem: continual WSI classification without storing previous slides or replay buffers. The method is coherent and reasonably complete, and all reviewers agree that the submission is above or near the acceptance threshold. The reported performance and storage efficiency are promising, and the inclusion of both task-CL and class-CL settings strengthens the evaluation.

    Enhance for final version paper: The authors should clarify the scalability of the growing KMP, including whether adaptive memory allocation, merging, pruning, or compression is possible. They should provide or at least discuss sensitivity analysis for key hyperparameters, especially memory size M, retrieval size K, and the loss weights. They should justify the max-pooling query design and report whether retrieved Top-K memory units correspond to the correct task in class-CL. Additional visualizations of retrieved memory units, feature shifts before and after PFC, and prototype changes after TC would directly address interpretability concerns. The authors should also clarify metric consistency across task-CL and class-CL, fix notation issues, and strengthen discussion of why the empirical gains are meaningful despite modest margins.



back to top