Abstract

Multiple instance learning (MIL) has become a standard paradigm for whole slide image (WSI) analysis in digital pathology, as it enables slide-level prediction without dense annotations. Existing MIL methods typically rely on exhaustive extraction and encoding of high-resolution patches. However, this practice suffers from two critical limitations in real-world clinical settings: it struggles to capture global visual cues at lower magnifications, and incurs substantial computational overhead due to the massive number of high-resolution patches per slide. To address these limitations, we propose an efficient low-resolution multiple instance learning (LRMIL) framework that transfers high-resolution knowledge to low-resolution representations. LRMIL adopts a two-stage distillation strategy. First, a patch-level cross-resolution distillation aligns low-resolution patch embeddings with high-resolution representations. Second, a slide-level knowledge distillation trains a low-resolution student MIL model under both slide-level supervision and teacher guidance. At inference time, LRMIL operates exclusively on low-resolution patches, substantially reducing data preprocessing and computational cost. Extensive experiments on multiple WSI benchmarks demonstrate that LRMIL consistently outperforms state-of-the-art MIL methods while achieving more efficient inference. These results highlight LRMIL as a practical and scalable solution for WSI analysis in clinical pathology.

Links to Paper and Supplementary Materials

Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/6341_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{ShiYon_LRMIL_MICCAI2026,
        author = { Shin, Yonghan AND Jeong, Won-Ki},
        title = { { LRMIL: Efficient Low-Resolution Multiple Instance Learning via High-Resolution Knowledge Distillation for Whole Slide Image Classification } },
        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

    Existing MIL methods typically rely on exhaustive extraction and encoding of large-scale patches from high-resolution (HR) WSIs, which limit the capture of global visual features in low-resolution (LR) WSIs and cause high computational cost. In this manuscript, the authors proposed a novel framework, LRMIL, based on two-stage knowledge distillation to decouple training and inference resolutions. The main contributions can be summarized as follows:

    • Patch-level Cross-resolution Distillation. This module is devised for transferring fine-grained semantic features from HR patches to LR features at the patch level, independently of MIL.
    • Slide-level Knowledge Distillation. Based on the LR patch features from the former module, an attention-based MIL model that performs accurate bag-level prediction using only LR patches is trained in this module.
  • 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 is well motivated and easy to follow. The major problem is clearly defined, and the proposed method is easy to follow. In particular, the idea of two-stage distillation is reasonable to solve the LR patches inference problem.

    • The proposed LRMIL method is meaningful in real-world pathology workflows, since it allows the decoupling of training (HR patches) and inference resolutions (LR patches). This will be helpful for establishing computionally efficient pathology applications.

    • Extensive experiments on diverse datasets and tasks are done to demonstrate that the proposed LRMIL achieves superior performance and computational efficiency than other SOTA methods.

  • 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 evaluation of molecular classification is somewhat not sufficient. Since the task is formulated as molecular subtype classification, the results would be more informative if the performance for predicting each subtype, rather than only overall accuracy.

    • The formulation of “survival prediction” in this work is not fully aligned with survival analysis in computational pathology. In this manuscript, this task is formulated as classification of survival status (alive vs. dead), evaluated by accuracy and AUC. This should be defined as survival status classification rather than survival prediction. A standard survival prediction should consider censoring, patient stratification, and ranking-based metrics such as the C-index.

    • More technical details are needed for the evaluation of otal inference time. The manuscript does not clearly describe the details about this experiment, including the time breakdown across patch extraction, feature encoding, and MIL aggregation, as well as relevant details such as WSI resolution, and preprocessing steps.

  • 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 main strengths of this work are its meaningful motivation, clear methodology design, and extensive experimental results. The method is not the most conceptually novel among the submissions in my batch, but the proposed LRMIL demonstrates its potential effectiveness in LR patch-based model inference, which is helpful for downstream applications.

    Consequently, if the authors address the major concerns in the experiments, I will give it a weak acceptance.

  • 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

    LRMIL addresses the computational inefficiencies and bottlenecks caused by high-resolution (HR) images in Multiple Instance Learning (MIL)-based Whole Slide Image (WSI) analysis. The manuscript proposes a teacher-student knowledge distillation framework designed to achieve performance comparable to HR-based models using only low-resolution (LR) inputs. The proposed methodology aligns information between LR and HR patches, which differ by a factor of 4 in magnification (a $4^2 = 16$ fold difference in patch count). This alignment is further extended to the slide and instance levels during the Stage 2 training phase. Additionally, a pseudo-label-based loss is introduced to refine the top-k attention alignment with greater precision. The experimental results provide a plausible and acceptable level of interpretability.

  • 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 definition is clearly articulated. The manuscript maintains a consistent narrative from the Abstract to the Conclusion, supported by relevant references and a well-structured methodology.
    • The proposed approach of aligning averaged LR token-level information with corresponding HR sub-region patches is highly novel. This effectively overcomes the inherent limitation where LR images contain coarse global context while HR patches provide the necessary fine-grained details.
    • The soft alignment mechanism goes beyond conventional teacher-student distillation by not being limited solely to the output layer. This design leverages the fact that most modern MIL architectures are attention-based, allowing for deeper feature-level knowledge transfer.
  • 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 is a significant point of confusion in the interpretation of Equations (1) and (2). The index $K=16$ representing the corresponding HR patches and the number of sub-tokens within a single LR ViT token (also 16) happen to coincide. Although these represent distinct components, their numerical equivalence complicates the reader’s understanding. Could the manuscript provide a more distinct notation or add a brief explanation to clearly decouple these two elements?
    • The training process of LRMIL appears to be discrete between Stages 1 and 2.This raises a slight concern regarding whether such a decoupled approach can consistently guarantee optimal performance compared to an end-to-end training scheme.
    • There is insufficient information regarding the selection of hyperparameters such as $T_1$, $T_2$, and the $\lambda$ values. The manuscript lacks a clear justification for the specific values used and the rationale behind their selection.
  • 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.

    (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?
    • Based on the contributions and strengths outlined above, I recommend an Accept for this manuscript.
    • The manuscript tackles a long-standing challenge in the MIL and proposes a highly novel solution.
  • 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

    The paper proposes LRMIL, a WSI MIL framework that transfers high-resolution knowledge to a low-resolution model through two-stage cross-resolution distillation. Its practical contribution is enabling inference with only low-resolution patches, with the goal of reducing the cost of conventional high-resolution MIL pipelines while maintaining competitive performance.

  • 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 practically important problem in computational pathology, namely how to reduce the high cost of high-resolution MIL for WSI analysis without sacrificing predictive performance. The final model uses only low-resolution patches, which is more attractive for scalable WSI analysis than methods that still depend on high-resolution processing at test time. This makes the work relevant to real deployment settings rather than only benchmark optimization. 2.Another strength is the teacher–student design, which makes low-resolution inference feasible through cross-resolution distillation. The work emphasizes the trade-off between performance and efficiency, and provides reasonably comprehensive validation across datasets, tasks, and ablations.

  • 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 methodological novelty appears moderate rather than strong. The paper combines several existing ingredients—attention-based MIL, teacher–student distillation, bag-level KD, and instance-level supervision—in a cross-resolution setting. While the integration is reasonable, the novelty seems to lie more in the formulation and combination of existing components than in a fundamentally new algorithmic idea. 2.Some empirical claims are stronger than what the results fully support. The paper states that LRMIL consistently outperforms prior methods, but the tables show a more mixed picture, with some baselines remaining competitive or stronger on particular datasets or metrics. Therefore, conclusions should be stated more carefully. 3.The survival prediction setting is somewhat simplified. Survival is formulated as a binary alive-vs-dead classification problem, which is less standard and less informative than time-to-event survival modeling with censoring. This limits the clinical relevance of the survival experiments. 4.Some implementation details could be clearer for full reproducibility. While the overall framework is understandable, the paper could better specify practical details.

  • 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?

    I gave a weak accept because the paper studies a practically relevant problem in WSI analysis and presents a reasonably coherent teacher–student framework for enabling low-resolution inference through cross-resolution distillation. The paper is generally well aligned with its motivation. At the same time, I have several reservations. The methodological novelty is somewhat limited, as the work mainly combines existing MIL and distillation components in a cross-resolution setting rather than introducing a strongly new algorithmic idea. In addition, some aspects of the experimental design and evaluation could be stronger, and a few of the paper’s claims are stated somewhat too strongly relative to the reported results. Overall, I found the work useful and practically relevant enough for a positive recommendation, but with clear limitations in novelty and experimental rigor.

  • 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.

    This paper proposes an efficient low-resolution MIL framework that transfers high-resolution knowledge to a low-resolution WSI model through two-stage cross-resolution distillation. The reviewers find the problem well-motivated and practically important. They also appreciate the clear teacher-student design and the experimental evaluation across datasets and tasks. The main concerns are relatively limited, including more technical details, clearer notation, hyperparameter and experiment settings, etc. The authors are encouraged to clarify the raised comments and refine the corresponding claims in the final version.



back to top