List of Papers Browse by Subject Areas Author List
Abstract
Digital histopathology images are fundamental to modern oncology, enabling precise cancer diagnosis and the identification of clinically relevant morphological features. Due to the gigapixel scale of Whole Slide Images (WSIs), the dominant paradigm remains a two-step pipeline that encodes patches with a pre-trained backbone and aggregates features via Multi Instance Learning (MIL). However, this paradigm inherently limits optimization; encoders are often kept fixed, and prototype-based approaches are not always tightly coupled with the slide-level discriminative objective. Here, we propose Marginal-Constrained Morphological Prototype Learning (MMPL), a framework that formulates WSI diagnosis as the selective retrieval of informative patches. MMPL introduces two core innovations: (1) global prototype learning optimized via Sinkhorn-Knopp assignment against a momentum-stabilized feature queue (2) prototype-guided top-k patch selection. By focusing only on the top-k most informative patches, MMPL reduces the number of processed instances up to 99.3%, substantially lowering computational requirements. Experimental results on three WSI benchmarks demonstrate that MMPL outperforms the previous best baseline by up to 4.2% in weighted F1 for metastasis detection and 1.8% for tumor subtyping. Crucially, this breakthrough efficiency enables the vision encoder to learn discriminative morphological features, further rendering end-to-end training practical. The code will be available at https://github.com/dmis-lab/MMPL.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/6616_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to the Code Repository
https://github.com/dmis-lab/MMPL
Link to the Dataset(s)
N/A
BibTex
@InProceedings{ParSih_Marginal_MICCAI2026,
author = { Park, Sihyeon AND Park, Jungwoo AND Kim, Hyunjae AND Kang, Jaewoo AND Kim, Bumsoo},
title = { { Marginal Constrained Morphological Prototype Learning for Patch Search in Whole Slide 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
The paper proposes MMPL, a prototype-based WSI framework to improve patch selection and computational efficiency in slide-level prediction. The method learns global morphological prototypes using a queue and Sinkhorn-Knopp-based balanced assignment, and then uses these prototypes to retrieve top-k informative patches for slide classification. Evaluation results on three WSI benchmarks are presented to demonstrate the advantages of the proposed method.
- 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 an important problem in WSI analysis: how to make patch selection more targeted rather than typical all-patch MILs. 2.The paper includes several useful analyses, including low-tumor-prevalence robustness, prototype assignment visualization, and ablations on the queue, prototypes, and clustering strategy. 3.Some of the results look strong, such as the performance gain on CAMELYON16 and TCGA-NSCLC with the ResNet50 backbone.
- 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 method description is too short and hard to understand. The paper states that prototypes retrieve top-k relevant patches, but it does not clearly explain the exact rules. How many patches are selected per prototype, whether duplicates are allowed across prototypes, and how the final subset is assembled before aggregation? 2.The efficiency argument also feels incomplete. The paper emphasizes processing fewer instances, but it still seems that all patches need to be encoded at least once to populate the queue and compute prototype similarities. Without runtime or memory measurements, it is hard to know what the practical gain really is. 3.The UNI-backbone results look inconsistent. On CAMELYON16, the gain is very large, whereas on TCGA-NSCLC, MMPL actually underperforms PANTHER on AUC. The results need more explanation.
- Please rate the clarity and organization of this paper
Poor
- Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.
The authors claimed to release the source code and/or dataset upon acceptance of the submission.
- Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?
N/A
- Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2026/en/REVIEWER-GUIDELINES.html
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 has a solid underlying idea and some encouraging results, but the current presentation and validations are not very clear and not totally convincing, as stated in the Weaknesses above.
- 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 rebuttal clears up some of the confusion, especially on the retrieval rule, the UNI result, and where the efficiency gain is supposed to come from. Still, I do not think it is enough to support acceptance. The method still feels only moderately novel, and the current validation is not sufficient to clearly separate the method from related work. I keep my score at Weak Reject and my final recommendation is Reject.
Review #2
- Please describe the contribution of the paper
This paper proposes MMPL, a prototype-driven MIL framework for WSI analysis that integrates dynamic global prototype learning and optimal-transport-based assignment to guide top-k patch selection. The key idea is to align prototype learning with slide-level supervision and enable efficient end-to-end training by focusing only on informative patches.
- 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 combination of momentum-based prototype updates and OT-based assignment is reasonable and well integrated. Compared to prior prototype methods (e.g., PANTHER), the idea of aligning prototypes with task objectives is meaningful. 2.The top-k selection strategy significantly reduces computation while maintaining or improving accuracy, which is important for real-world WSI deployment. 3.The method shows consistent improvements across multiple WSI benchmarks, with clear gains in WF1 and AUC. Ablation studies are also reasonably thorough.
- 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 proposed prototype-based top-k patch retrieval is quite similar in spirit to recent MoE-style routing methods that also select a subset of informative patches [1]. However, the paper does not clearly explain how this approach differs from or improves over these methods. In particular, it is unclear whether the benefit comes from the prototype design itself or simply from sparse patch selection. This makes the positioning of the method less clear. 2.Some parts of the method (e.g., interaction between queue, momentum encoder, and OT assignment) are not explained as clearly as they could be, making the framework slightly harder to follow. 3.While some qualitative results are shown (Fig. 3), there is still limited quantitative analysis of how prototypes evolve, how stable they are, and how they relate to pathological semantics across datasets. [1] Wu, Junxian, et al. “Learning heterogeneous tissues with mixture of experts for gigapixel whole slide images.” Proceedings of the Computer Vision and Pattern Recognition Conference. 2025.
- 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?
Overall, the paper is technically solid and shows promising results, but there are still some important issues that need to be addressed. In particular, the positioning with respect to recent MoE-based patch selection methods is unclear, and several key design choices (e.g., prototype learning mechanism) are not sufficiently analyzed or explained. These make it harder to fully understand where the performance gains come from. That said, I think the idea is interesting and relevant to the field. If the authors can better clarify the differences from MoE-style methods and provide a deeper analysis of the prototype behavior, I would be happy to raise my score.
- 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.
My main concerns were resolved during the rebuttal phase, so I recommend acceptance after the rebuttal.
Review #3
- Please describe the contribution of the paper
The paper proposes a prototype-driven framework for WSI analysis that formulates slide-level prediction as a selective patch retrieval problem. Its main contributions include dynamically learning global morphological prototypes via a momentum-updated feature queue, introducing an optimal transport–based assignment (via Sinkhorn-Knopp) to enable balanced and task-aware prototype matching, and performing BoVW-style weighted aggregation over top-k selected patches to improve both computational efficiency and representation quality. Together, these designs allow more effective identification of informative regions while supporting end-to-end training of the encoder.
- 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 has several notable strengths: it presents a well-motivated and coherent framework that integrates prototype learning, optimal transport, and top-k patch selection into a unified pipeline; the use of Sinkhorn-based assignment provides a principled way to enforce balanced and task-aware prototype utilization; and the dynamic prototype learning with a momentum queue is technically sound and helps capture global morphological patterns. In addition, the method demonstrates strong empirical performance across multiple benchmarks, while significantly improving computational efficiency by focusing only on informative patches, making the approach both effective and practical.
- 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 main limitation of the paper lies in its relatively incremental novelty, as the overall framework shares notable similarities with prior prototype-based and selective aggregation methods for WSI analysis. Although the use of optimal transport and top-k retrieval introduces some technical differences, the conceptual advancement over existing approaches is not entirely clear. More importantly, the paper does not include direct comparisons with closely related work (e.g., methods based on dynamic encoding or prototype-driven representation learning [1]), making it difficult to fairly assess the actual performance gains. Additionally, some design choices (such as the specific benefits of OT assignment over simpler alternatives) are not sufficiently justified through ablation or analysis.
[1] Dynamic Residual Encoding with Slide-Level Contrastive Learning for End-to-End Whole Slide Image Representation.
- 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.
(6) Strong Accept — must be accepted due to excellence
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
I lean toward accepting this paper due to its solid technical design and practical effectiveness. While the overall idea is somewhat related to existing prototype-based and selective aggregation methods, the integration of momentum-based global prototype learning with optimal transport assignment is well-executed and provides a principled way to improve prototype utilization. The framework is coherent, and the top-k retrieval strategy offers clear computational benefits without sacrificing performance. In addition, the experimental results are strong and consistently demonstrate improvements across multiple benchmarks. Despite some missing comparisons with closely related work, the paper still provides a meaningful and practically valuable contribution to WSI analysis.
- Reviewer confidence
Very confident (4)
- [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 maintain my recommendation for acceptance.
Author Feedback
We sincerely thank all reviewers for their thoughtful and constructive feedback. Below, we provide point-by-point responses. (Reviewer 1)
- Patch selection rules: We compute the cosine similarity between the patch features of the input WSI and the learned prototypes. The total retrieval budget of k patches is then dynamically partitioned among prototypes in proportion to their slide-level attention weights, which are learned via a gated attention layer trained on the downstream task. Each prototype retrieves its assigned share of the most similar patches, and the retrieved instances are concatenated into the final informative subset capped at k. Since the OT formulation via Sinkhorn-knopp promotes uniform assignments, the retrieved patches are encouraged to align with a single prototype, although minor overlaps may occur as this is not enforced explicitly.
- Inconsistent UNI‑backbone results: The WF1 and AUC results for MMPL with UNI on the TCGA-NSCLC dataset were incorrectly reported. The correct performance is WF1: 0.950 ± 0.03 and AUC: 0.980 ± 0.003 (originally misreported as WF1: 0.984 and AUC: 0.907). We apologize for any confusion. Upon careful re-examination of all logs and result tables, we confirmed that this was the only reporting error. Importantly, despite this correction, the main conclusions remain unchanged. With the UNI backbone, MMPL still outperforms PANTHER on most metrics, while showing competitive AUC on TCGA-NSCLC.
- Efficiency: While all patches are encoded during the forward pass, the primary memory bottleneck lies in storing intermediate activations for backpropagation during training. Our efficiency gain stems from selective backpropagation. In fact, training ABMIL with 10,000 patches using a ResNet50 is infeasible (1.06TB of memory). In contrast, MMPL restricts gradient updates to only the top-100 patches (only 22.26GB), enabling end-to-end optimization even with a single NVIDIA RTX 3090.(Reviewer 2)
- Distinctions from MoE-style routing (PAMoE): While MMPL and PAMoE share identifying informative patches, a key difference is reliance on external priors for expert specialization. MoE-based models typically learn gating via a weighted sum of diverse expert outputs through downstream loss. However, in the WSI-MIL setting, expert specialization is often difficult to achieve due to data scarcity and low label complexity, which can lead to unstable routing. PAMoE also acknowledges this challenge in their paper (see Section 4.2). In contrast, MMPL formulates self-supervised learning via an OT-based framework. Using a shared MLP across all patches and Sinkhorn-knopp, our prototypes capture diverse morphological patterns without requiring external knowledge. Thus, the improvements stem from prototype design rather than simple patch selection.
- “Interaction between the queue …”: By employing an EMA-updated encoder for stable representation shifts, we populate the feature queue to mitigate the local bias inherent in individual slides. Within this stabilized global context, we formulate prototype assignment as an OT problem with Sinkhorn-Knopp. This enables robust prototype evolution from scratch during training, which then guides a dynamic top-k retrieval process.
- Quantitative Prototype analysis: While we agree that a more comprehensive quantitative analysis would further strengthen the paper, additional experiments are difficult to include due to the rebuttal guidelines regarding new results. Importantly, the current results in Table 2 and Fig 3c provide quantitative evidence that the prototypes avoid collapse and capture relevant pathological semantics on CAMELYON16 with high quantitative consistency. (Reviewer 3)
- Novelty over prior work (DRE-SLCL): While DRE-SLCL relies on random sampling and external reports, MMPL learns global prototypes in a self-supervised manner during training. This enables the vision encoder to be optimized on informative patches retrieved through prototype guidance.
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 prototype-driven MIL framework for WSI analysis that integrates dynamic global prototype learning with top‑k patch selection to improve both computational efficiency and slide‑level classification performance. The reviewers agree that the method shows solid and consistent improvements on several WSI datasets. The weaknesses include insufficiently clear method description and lack of comparison with closely related work. Please respond to the major concerns of the reviewers, including the following points: clarify the exact patch selection rules,explain the inconsistent UNI‑backbone results,clarify the difference from MoE‑style routing methods,explain the interaction between the queue, momentum encoder, and OT assignment.
- 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.
This paper received two rejects and one accept after rebuttal. I recommend rejection of this paper and the main reason is the insufficient evaluation. The major results are reported with ResNet50 as the backbone, but its performance is much lower than using UNI as the backbone, which makes the Resnet50 results useless in practice. As a pathology foundation model (such as UNI) can result in higher performance, only improvement over this strong baseline is practically meaningful.
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 paper presents a well-motivated and practically relevant framework for efficient WSI analysis through prototype-driven patch selection. Reviewers appreciated the integration of dynamic prototype learning, optimal-transport-based assignment, and top-k patch retrieval within a unified framework, as well as the strong empirical performance across multiple benchmarks. The method also demonstrates meaningful computational advantages by focusing on informative patches while maintaining competitive accuracy. Although concerns were raised regarding novelty and additional comparisons with related prototype-based methods, the rebuttal addressed many of these issues satisfactorily. Overall, the paper provides a solid contribution to efficient and scalable WSI analysis, and I support acceptance.
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.
This is an interesting piece of work, and the authors have satisfactorily addressed the majority of the reviewers’ concerns during the rebuttal.
