Abstract

The tumor region plays a key role in pathological diagnosis. Tumor tissues are highly similar to precancerous lesions and non-tumor instances often greatly exceed tumor instances in whole-slide images (WSIs). These issues cause instance-semantic entanglement in multi-instance learning frameworks, degrading both model representation capability and interpretability. To address this, we propose an end-to-end prototype instance-semantic disentanglement framework with low-rank regularized subspace clustering, namely PID-LRSC, in two aspects. First, we use secondary instance-subspace learning to construct low-rank regularized subspace clustering (LRSC), addressing instance entanglement caused by an excessive proportion of non-tumor instances. Second, we employ enhanced contrastive learning to design prototype instance‒semantic disentanglement (PID), resolving semantic entanglement caused by the high similarity between tumor and precancerous tissues. We conduct extensive experiments on multicentre pathology datasets, implying that PID-LRSC outperforms other state-of-the-art methods. Overall, PID-LRSC provides clearer instance semantics during decision-making and significantly enhances the reliability of auxiliary diagnostic outcomes. Code and datasets are available at https://github.com/Prince-Lee-PathAI/PID-LRSC.

Links to Paper and Supplementary Materials

Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/0426_paper.pdf

SharedIt Link: Not yet available

SpringerLink (DOI): Not yet available

Supplementary Material: Not Submitted

Link to the Code Repository

https://github.com/Prince-Lee-PathAI/PID-LRSC

Link to the Dataset(s)

AMU-LSCC: https://pan.baidu.com/s/1ULxv4MCbv0TUz58UXSv29A?pwd=RYES AMU-CSCC: https://pan.baidu.com/s/1x64wZWFfXgKeRc44hq5atw?pwd=8LR6 DHMC-Kidney: https://pan.baidu.com/s/1Ekpx81dyJczsteTfnVNG6g?pwd=N74z DHMC-Lung: https://pan.baidu.com/s/1Au1Qr9kN8T6YPK9mFlTPBg?pwd=65o8

BibTex

@InProceedings{LiChe_Prototype_MICCAI2026,
        author = { Li, Chentao AND Huang, Pan},
        title = { { Prototype Instance-Semantic Disentanglement with Low-Rank Regularized Subspace Clustering for WSIs Explainable Recognition } },
        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 PID-LRSC, a MIL framework for WSI classification that aims to disentangle instance and semantic entanglement between sparse tumor regions and abundant non-tumor/background patches. Experiments on two private SCC datasets and two public datasets show consistent improvements over recent MIL baselines, with visualizations suggesting clearer instance-level 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.

    1.The low-rank metric-learning scheme is a simple but effective way to bias clustering toward discriminative, low-redundancy subspaces under extreme instance imbalance. 2.The integration of prototype-based semantic anchoring with subspace clustering addresses both instance-level redundancy and tumor/precancerous ambiguity within a single training loop. 3.Addressing sparse tumor signal dilution and semantic confusion with precancerous regions is highly relevant to real-world WSI use and clinical interpretability. Broad comparisons against strong MIL baselines on four datasets, including two from an external center, show consistent gains in ACC 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.Inconsistent naming (CDF vs CFD), and some notational/typographical glitches that make sections harder to parse. The supplementary documents lack crucial information, such as a description of the private dataset. 2.The current evaluation provides a limited ablation analysis; specifically, it lacks a detailed investigation into how the model’s performance fluctuates with changes to $r$, $\gamma_{1/2/3}$, the number of clusters, or specific prototype configurations. 3.The success of PID critically depends on the prototypes, but their provenance, number, selection, and maintenance are insufficiently specified; robustness to prototype noise is not studied. 4.Although Table 1 specifies the use of ‘Swin-T pretrained features’, it remains unclear whether this ViT encoder is fine-tuned during the end-to-end training process or strictly frozen as an offline feature extractor. This ambiguity regarding the exact source of the trainable parameters, combined with a lack of comprehensive training protocol details, significantly hampers the reproducibility of the proposed method.

  • 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 has provided an anonymized link to the source code, dataset, or any other dependencies.

  • 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?

    The research motivation is clear and well-articulated, and the experimental results across the four datasets are highly encouraging. However, there are a few notable shortcomings regarding the presentation of the datasets and implementation details, which currently make it difficult to fully assess the rigor and reproducibility of the study. Additionally, the description of the prototype construction and its robustness is insufficient. If the authors can enhance the clarity of the methodology (particularly concerning the handling and formulation of the prototypes) and expand upon the ablation and sensitivity analyses, this work will represent a significant contribution to the field.

  • 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.

    I have been satisfied with authors’ rebuttal responses.



Review #2

  • Please describe the contribution of the paper

    This paper proposes an interpretable WSI classification framework, PID-LRSC, that aims to address what the authors call “instance-semantic entanglement” in MIL-based pathology analysis. The method combines two main ingredients. First, it introduces a two-phase low-rank regularized subspace clustering module to separate sparse tumor-related instances from abundant non-tumor/background regions in a learned subspace. Second, it introduces a prototype-guided disentanglement module that uses prototype-to-subspace discrepancy to assign semantic roles to clustered groups and refine the final WSI representation. The paper further claims that this design improves both classification performance and interpretability on multicentre pathology datasets. These central claims and the two-stage decomposition of the method are described in the main submission.

  • 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 tackles a genuinely important and clinically relevant problem in computational pathology. Specifically, it addresses a well-known weakness of multiple instance learning (MIL) approaches applied to whole-slide images (WSIs): the extreme sparsity of diagnostically informative tumor regions amidst an overwhelming majority of non-tumor tissue, compounded by the semantic ambiguity between malignant and precancerous regions under weak supervision. The authors’ framing of this challenge as an “entanglement” problem—both at the instance and semantic levels—is intuitive, well-motivated, and particularly pertinent for developing explainable pathology models. This motivation is clearly articulated in the introduction and effectively illustrated in Figure 1 of the main paper.

    Notably, the proposed method attempts to jointly advance representation learning and interpretability, rather than treating interpretability as a post hoc add-on. Instead of merely generating attention-based heatmaps after the fact, the authors seek to inject structural priors directly into the learned representation through clustering and prototype-guided refinement. This constitutes a more principled design choice compared to standard attention visualization strategies commonly adopted in the WSI literature.

    The low-rank metric perspective introduced in the method is potentially interesting. In particular, the use of a learned positive semi-definite metric with enforced low-rank structure to guide subspace clustering offers a reasonable mechanism for suppressing redundant non-tumor variation. The coarse-to-fine formulation is conceptually sound, although certain technical aspects—such as the precise construction of prototype instance subspaces and the sensitivity to the fixed number of clusters—remain underspecified.

    The empirical evaluation appears reasonably comprehensive at first glance. The authors compare their approach against a substantial set of recent WSI MIL baselines and report both accuracy (ACC) and area under the curve (AUC) across multiple datasets. The inclusion of ablation studies (Tables 1–3) and qualitative visualizations further suggests a concerted effort to substantiate their claims beyond performance on a single benchmark.

    Importantly, interpretability is treated as a first-class objective—not just as a secondary benefit. The manuscript goes beyond claiming improved classification performance and actively attempts to align the disentangled, model-identified regions with pathologist-annotated ground truth boundaries. While this alignment is not yet fully quantified with rigorous metrics, the fact that the authors are explicitly engaging with this critical question sets their work apart from many WSI-focused papers that overlook clinical interpretability altogether.

  • Please list the major weaknesses of the paper. Please provide details: for instance, if you state that a formulation, way of using data, demonstration of clinical feasibility, or application is not novel, then you must provide specific references to prior work.

    The paper is built around an intuitively appealing idea, but in its current form, the main weakness is that several of its most important design choices are asserted rather than convincingly established. In particular, the decision to fix the clustering structure at three groups is insufficiently justified. Because this choice sits at the core of the proposed semantic decomposition, it cannot be treated as a minor implementation detail. The paper offers only a brief heuristic explanation: tumor, non-tumor, and background are common categories in WSI analysis. This does not amount to a principled argument that three clusters are appropriate for all datasets studied, nor does it show that performance is robust to alternative cluster numbers. As a result, the method risks looking partially tailored to a presupposed semantic partition rather than learning that structure from data. This concern is reinforced by the lack of a stronger sensitivity analysis around this design decision.

    A second major weakness is that the prototype component, which is essential to the paper’s semantic disentanglement story, is not described with enough precision. The method repeatedly relies on a prototype instance subspace and even uses it directly in the loss formulation and semantic assignment process, yet the manuscript never makes fully clear how these prototypes are obtained, initialized, updated, or supervised. From the figures and wording, one can infer that domain knowledge may be involved, but it remains ambiguous whether the prototypes are externally provided, weakly selected, learned jointly, or derived from some additional annotations. This is not a minor omission: without a transparent description of prototype construction, it becomes difficult to assess whether the method is truly reproducible and whether it operates under the same supervision assumptions as the baselines. In a paper whose central claim is that prototypes resolve semantic ambiguity, the provenance of those prototypes should be explicit and technically complete.

    Another serious issue is the inconsistency between the main submission and the supplementary material. The main paper presents PID-LRSC, whereas the uploaded supplementary file appears to describe a different method, with a different title, different terminology, different equations, and even a different dataset configuration. At present, the relationship between these two documents is not explained. This creates a practical problem for the reviewer because some details missing from the main paper are said to be deferred to the supplement, yet the supplement does not read as a direct extension of the submitted manuscript. More importantly, it creates a procedural concern: I could not determine whether this was an upload mistake, an earlier version of the work, or a separate, related submission that should have been explicitly contextualized. In any case, it weakens confidence in the submission package and substantially limits the paper’s reproducibility.

    The paper is also somewhat weaker than it appears at first glance in the alignment between its conceptual narrative and its actual formulation. For example, the manuscript describes the method as involving enhanced contrastive learning, but the technical section mainly develops a discrepancy-based comparison between clustered subspaces and prototype representations, followed by rank-based semantic assignment and refinement. I did not find a clearly defined contrastive objective that matches the framing in the introduction. More broadly, the paper often uses strong interpretive language about disentanglement, causality, and semantic clarity, but the underlying optimization remains closer to a structured clustering-and-reweighting pipeline than a rigorously demonstrated disentanglement framework. This does not mean the method is ineffective, but it does mean that some of the claims are currently stronger than the evidence provided.

    Finally, while the empirical results are promising, the interpretability claims are still supported more by selected qualitative examples than by a sufficiently rigorous quantitative evaluation. The visualization results are suggestive, and the alignment with annotated regions is encouraging, but for a paper that places explainability at the center of its contribution, I would have expected stronger evidence, such as systematic localization metrics, agreement analysis, or a more structured clinical validation of the highlighted regions. The writing and presentation also need polishing, as inconsistencies in notation, formatting, and phrasing occasionally make it harder than necessary to follow the technical argument. Taken together, these issues do not erase the potential of the work, but they do make the current version feel methodologically under-specified and not yet fully mature for acceptance.

  • 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 has provided an anonymized link to the source code, dataset, or any other dependencies.

  • 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?

    Although I do not have enough information to conclude that there is a definite ethics violation, I note a procedural/research-integrity concern regarding the supplementary material. The uploaded supplement appears to contain a separate full manuscript with a different title, method description, and experimental setup, rather than a conventional appendix directly supporting the submitted paper. However, the submission does not explain the relationship between the main paper and this supplementary manuscript, nor does it clarify whether this is an earlier version, a related concurrent submission, or another work on the same topic. If this supplementary document is indeed a different paper on a closely related topic, the authors should explicitly disclose and justify the distinction. In its current form, the lack of such clarification creates uncertainty about submission consistency and compliance with conference policy.

  • 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

    The paper presents a compelling intuition: in WSI MIL, both weak performance and limited interpretability may originate from a fundamental mismatch between pathological structures and the latent organization induced by bag-level supervision. This represents a valuable and worthwhile perspective. However, the current version falls short of establishing a sufficiently complete “evidence chain” that connects the initial assumption through the proposed method to comprehensive validation.

    Two critical revisions are essential for strengthening the manuscript. First, the authors must clearly articulate how prototypes are obtained and specify the supervision mechanisms they rely on. Second, the inconsistency between the main manuscript and supplementary material requires resolution to ensure methodological coherence throughout the paper.

    Additionally, I would strongly encourage the authors to incorporate a sensitivity analysis of the number of clusters, ideally comparing the fixed k=3 setting against adaptive or larger-k configurations. Currently, the three-group decomposition appears more assumed than empirically demonstrated. Furthermore, the interpretability claims would benefit significantly from quantitative localization metrics or a more systematic reader study, rather than relying solely on qualitative examples, thereby substantially strengthening the paper’s contributions to the field.

  • 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?

    My assessment is shaped less by the headline results and more by the opportunity to strengthen the methodological exposition. The paper presents a promising idea that addresses an important challenge in WSI MIL, and the reported empirical gains are nontrivial and encouraging. The core intuition—that weak performance and interpretability may stem from a mismatch between pathological structure and bag-level supervision—is insightful and warrants further exploration.

    That said, several aspects would benefit from clarification to fully realize the paper’s potential. The source and learning mechanism of the prototypes could be articulated more explicitly, and the rationale behind the fixed three-cluster design would be strengthened with additional justification or sensitivity analysis. Similarly, clarifying the precise correspondence between the narrative claims and the optimization objective would enhance the methodological transparency.

    Despite these points for improvement, the work makes a valuable contribution to the field, and I believe it merits acceptance with the expectation that the authors will address these methodological clarifications in the final version. The foundational idea is sound, the empirical results are promising, and with modest refinements to the exposition, this could become a well-regarded contribution to WSI MIL research.

  • 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.

    After carefully reading the rebuttal and re-evaluating the submission, I am willing to support acceptance, although I still believe several aspects of the paper could be strengthened in the camera-ready version. My original concerns mainly focused on whether some of the core methodological choices were sufficiently justified and whether the paper provided enough technical transparency for reproducibility. The rebuttal substantially improved clarity on these issues. In particular, the authors clarified that the three-cluster design is intended as a coarse pathological prior rather than a claim about exact biological subtypes, which makes the formulation more reasonable in the context of WSI semantic decomposition. While I still think a broader sensitivity analysis over different values of K would further strengthen the work, the explanation that the groups correspond to lesion-related tissue, microenvironment/non-lesion tissue, and background/artifacts provides a coherent pathological interpretation that is consistent across datasets and aligns with common practice in WSI analysis. More importantly, the rebuttal clarified the prototype construction process in significantly greater detail. The authors now explicitly state that prototypes are annotated by five experts using unions of tumor regions and are only class-specific during training, with variability intentionally incorporated to improve robustness. This clarification addresses my earlier concern that the provenance and supervision assumptions behind the prototypes were under-specified. Although these details should ideally appear more explicitly in the main manuscript, the methodological pipeline is now sufficiently understandable and reproducible.



Review #3

  • Please describe the contribution of the paper

    This paper proposes PID-LRSC, an end-to-end multiple-instance learning framework for explainable whole-slide image recognition. The method targets two issues in WSI classification: instance entanglement caused by the dominance of non-tumor regions, and semantic entanglement caused by visual similarity between precancerous and tumor tissue.

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

    The paper addresses an important computational pathology problem: making MIL-based WSI recognition both accurate and interpretable. The motivation around sparse tumor regions, redundant non-tumor regions, and ambiguous precancerous/tumor morphology is clinically plausible.

    The proposed decomposition into instance-level subspace clustering and semantic prototype disentanglement is interesting. The method attempts to move beyond attention heatmaps by assigning instances to tumor, non-tumor, and background-like subspaces, which could improve interpretability if validated.

  • 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 method is difficult to fully understand and reproduce from the current manuscript. Many important dataset and implementation details are deferred to supplementary material, including dataset sizes, train/test splits, patch extraction, magnification, prototype construction, and exact feature-extractor settings. Since two key datasets are private, the public-dataset protocol and release details become especially important.

    The mathematical presentation has several ambiguities. The learned metric matrix dimensions are not always clear. The derivation that the nuclear norm equals the trace only holds under positive semidefinite assumptions, which should be stated carefully. The KMeans step is not differentiable, yet the paper repeatedly describes the full method as end-to-end; the actual gradient path through clustering should be clarified. The characteristic function discrepancy formulation is also underexplained, including how the sampling distribution and kernel are implemented.

    The role and construction of prototypes are insufficiently specified. The method relies on prototype instances as semantic anchors, but it is unclear how many prototypes are used, whether they require pathologist annotation, whether they are class-specific, whether they come from training slides only, and how sensitive the method is to prototype quality.

    The interpretability claims are promising but not quantitatively validated. The paper shows qualitative alignment with pathologist boundaries, but it does not report localization metrics against annotations, such as tumor-region Dice, pointing-game accuracy, overlap with annotated tumor regions, or correlation with expert ROI labels. Statistical reliability is not sufficiently reported. The tables give point estimates but no confidence intervals, standard deviations, or significance tests. This is important for WSI datasets, where slide counts may be limited and patient-level split variance can be high.

  • 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 tackles a relevant and challenging pathology problem and reports strong results against many MIL baselines. However, the current manuscript has significant clarity, reproducibility, and methodological-validation issues. The interpretability claims need stronger quantitative evidence, and the prototype/clustering mechanisms require clearer specification.

  • 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

For Meta Reviewer: Please kindly refer to R2.1 R3.1, R3.2 and R3.3 for your concerns.

Reviewer #1: R1.1 ‘CDF’ was a typo and has been revised. Critical details like private datasets are in abstract link because of double-blind. For Prototype construction, please kindly refer to the same response of R3.3.

R1.2 In Fig 2, PID-LRSC uses pretrained Swin-T fully open for end-to-end training. All compared models use frozen features extracted by Swin-T in two stages. They were trained by RMSprop with fixed random seed 0 for reproducibility. (See abstract link: section2 for details)

Reviewer #2: R2.1The choice K=3 is a coarse pathological prior rather than exactly three biological subtypes. The three groups correspond to factors: lesion/tumor-related tissue, microenvironment or non-lesion tissue, and background/artifacts such as blur, bubbles or staining artifacts.. This coarse partition is common across WSI analysis and generalizes across datasets. 

R2.2 For Prototype constructions, please kindly refer to the same response of R3.3. R2.3 The main paper and supplementary one are to comply with the MICCAI dual-submission policy for similar topics (see official guidelines).

R2.4 Enhanced contrastive learning is not formulated instance-wise but at subspace level. Specifically, c_min and c_max minimize/maximize the discrepancy between clustered subspaces and prototypes, achieving spatial-level disentanglement. And the final aggregation Z_WSI employs semantic-guided feature refinement with ranked prototype discrepancies, which alleviates semantic entanglement.

R2.5 Based on ground truth, heatmap and semantic clustering, we randomly selected WSIs to compute two average Dice scores. They are CSCC (0.4605, 0.4424), LSCC (0.7096, 0.2210), Lung (0.6198, 0.3341), and Kidney (0.6988, 0.4981).

Reviewer #3: R3.1 Critical details are in the abstract link. For example, two private datasets are CSCC (262 WSIs3) and LSCC (342 WSIs), split into the same 6:4 train/test ratio as public DHMC-Kidney and Lung; Patches per WSI are at 10x magnification; Prototype construction refers to R3.3; and ImageNet-pretrained Swin Transformer is used to extract frozen features for compared model training and fully open for end-to-end PID-LRSC training.

R3.2 (1) The learned matrix W is in shape [r,n]. Here n=768 is the output of ViT encoder and r=384 is a trade-off between efficiency and effectiveness. Matrix (W^T)W leads to a positive semi-definite matrix A. It completes the nuclear norm derivation. (2) Although K-means is non-differentiable, c_min, c_max and trace(A) are computed in the low-rank subspace parameterized by W, whose printed gradients are showed in the computational graph. So it can be propagated through via ‘autograd’. Thus, PID-LRSC enables end-to-end deep-clustering paradigm rather than isolated clustering. (3) For CFD: Given feature input Z and sampled frequency t=4096, define Phi_z(t)=E[cos(t^T Z)+isin(t^T Z)]. Then we can compute the phase function and amplitude function. See details and codes in abstract link.

R3.3 Prototype are given by 5 experts with double-blind using the union of tumor regions. 64 prototypes per WSI are included in CSCC , 81 for LSCC and 5 for public datasets. The prototype instances are class-specific only for training stage. Because of the multi-expert unions, they have large annotation variability to reduce subjectivity and improve training robustness. Also, enough quantity of prototype from different pathologists alleviates sensitivity to prototype quality. (Data and details see abstract link: section2)

R3.4 Based on ground truth, heatmap and semantic clustering, we randomly selected WSIs to compute two average Dice scores. They are CSCC (0.4605, 0.4424), LSCC (0.7096, 0.2210), Lung (0.6198, 0.3341), and Kidney (0.6988, 0.4981).

R3.5 We fixed seed=0 across all stages to ensure reproducibility and reduce variation caused by initialization and sampling. And our model focuses on slide-level representation capability.




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 MIL framework for WSI classification that combines low‑rank regularized subspace clustering to separate sparse tumor instances from abundant non‑tumor/background regions and prototype‑guided semantic disentanglement to resolve ambiguity between precancerous and malignant tissue. The reviewers agree that the paper addresses a well‑known challenge in computational pathology and show consistent and encouraging performance improvements (ACC/AUC) over a broad set of recent MIL baselines on four datasets. The reviewers raised question about the insufficient specification of prototype construction and learning, lack of clarity and details influencing the reproducibility, and oversimplified fixed clustering assumption (K=3) without justification. Please respond to the main concerns of the reviewers, including the following ones: How are prototypes obtained, initialized, updated, and supervised?What is the exact training setup regarding the feature extractor?How is the non‑differentiable K‑means step handled in the claimed “end‑to‑end” training?Provide a full description of the private datasets.

  • 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 receive an average score of 4, and one reviewer raised the score after rebuttal. Though the reviewer with negative score did not update the post-rebuttal score, most of his/her concerns have been addressed in the rebuttal. The author should incorporate the substantive informatoin in the rebuttal into the final version of the paper.



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.

    This paper was initially invited for rebuttal following the primary review period, where reviewers raised several concerns regarding insufficient specification of prototype construction and learning, and missing details. After carefully evaluating the authors’ rebuttal, the reviewers’ revised feedback, and re-examining the paper, I conclude that the authors have effectively addressed the key concerns raised, and the work now meets MICCAI’s publication standards.



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, the paper has received generally favorable reviews, with one reviewer not providing an update to their score. While some concerns with respect to the experimental setting have been answered, concerns on methodological choices, sensitivity analysis, and in-depth description regarding the provenance behind the prototype construction persist. Unless other concerns regarding the supplementary information in the code repository prevent acceptance, the paper is in my opinion slightly above the acceptance threshold.



back to top