Paper Info Reviews Author Feedback Meta-Review Back to top

List of Papers Browse by Subject Areas Author List

Abstract

Multiple instance learning (MIL) has been widely used for representing whole-slide pathology images. However, spatial, semantic, and decision entanglements among instances limit its representation and interpretability. To address these challenges, we propose a latent factor grouping-boosted cluster-reasoning instance disentangled learning framework for whole-slide image (WSI) interpretable representation in three phases. First, we introduce a novel positive semi-definite latent factor grouping that maps instances into a latent subspace, effectively mitigating spatial entanglement in MIL. To alleviate semantic entanglement, we employs instance probability counterfactual inference and optimization via cluster-reasoning instance disentangling. Finally, we employ a generalized linear weighted decision via instance effect re-weighting to address decision entanglement. Extensive experiments on multicentre datasets demonstrate that our model outperforms all state-of-the-art models. Moreover, it attains pathologist-aligned interpretability through disentangled representations and a transparent decision-making process. Code and datasets are available at https://github.com/Prince-Lee-PathAI/PG-CIDL.

Links to Paper and Supplementary Materials

Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/0366_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/PG-CIDL

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 CAMELYON16: https://pan.baidu.com/s/1JZ5HFIigPOQsaF_IY1T5WQ?pwd=143T DHMC-Kidney: https://pan.baidu.com/s/1Ekpx81dyJczsteTfnVNG6g?pwd=N74z

BibTex

@InProceedings{LiChe_Positive_MICCAI2026,
        author = { Li, Chentao AND Bozorgtabar, Behzad AND Ping, Yifang AND Wong, Chun-Ka AND Huang, Pan AND Qin, Jing},
        title = { { Positive Semi-definite Group-Aware Instance Disentangled Learning for WSI Representation } },
        booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
        year = {2026},
        publisher = {Springer Nature Switzerland},
        volume = {LNCS 16891},
        month = {September},
        page = {pending}
}

Reviews

Review #1

Please describe the contribution of the paper

This paper provides a framework for learning WSI representation. The author first analyzes the entanglements in spatial, semantic, and decision aspects. To address these issues, the author first utilizes a positive semi-definite matrix to learn an adaptive distance for clustering (Kmeans) . Then, they uses counterfactual inference to disentangle semantic meanings for different clusters. and finally applies instance effect re-weighting to aggregate WSI features for MIL. Experiments show the method improves classification 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.This paper analyzes three existing challenges in WSI representation learning and designs targeted solutions, leading to improved classification performance. 2.The paper integrates positive semi-definite metric learning with KMeans, making the instance clustering better match the real distribution of WSI features. 3.The paper compares many MIL methods and conducts experiments on four datasets.
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 authors submitted a supplementary material entitled Prototype Instance-semantic Disentanglement with Low-rank Regularized Subspace Clustering for WSIs Explainable Recognition, which appears unrelated to the main paper. Consequently, some critical information is missing, such as the description of the private dataset and the CID idea. 2.When instances from two different categories are mixed together due to K-means clustering error ,it is not clear how CID handles such clustering noise. 3.The method relies heavily on K-means clustering with a fixed K=3 (i. e. , Tumor, microenvironment, and background). This represents a oversimplification of the histological complexity in WSIs. Moreover, no sensitivity analysis is performed on the choice of K, which further limits this assumption. 4.Key ablation studies (Table2) are only conducted on a private dataset. Presenting ablation results on public datasets would greatly improve reproducibility.
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 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.

(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 overall score is based on several major concerns: The supplementary material is irrelevant and causes missing key information; the fixed K-means clustering oversimplifies tissue structure without sensitivity analysis and the stability of CID against K-means noise is unclear. Please see the weaknesses for detail.
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.

authors have cleared my concerns during the reubttal stage.

Review #2

Please describe the contribution of the paper

The paper proposes PG-CIDL, a three-phase disentangled representation learning framework for Multiple Instance Learning (MIL) based Whole Slide Image (WSI) analysis. The claimed contributions are: (1) Positive Semi-definite Latent Factor Grouping (PSD-LFG): a metric learning-based clustering approach using positive semi-definite matrices to group instances into three latent factors (tumor, microenvironment, background); (2) Cluster-reasoning Instance Disentangling (CID): a counterfactual inference method using KL divergence to measure causal effects of each group; and (3) Instance Effect Re-weighting: a generalized linear weighted aggregation based on measured effects to mitigate decision entanglement.
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 correctly identifies three types of entanglement (spatial, semantic, decision) in existing MIL frameworks, which is a valid and clinically relevant concern. Experiments show the superior performance on four datasets, with significant improvements over baselines.
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.

K-Means is Not Differentiable: The paper acknowledges this but proposes a “group-separation regularizer” d_reg as a workaround. However, this regularizer only encourages separation between tumor and non-tumor means. It does not actually optimize the clustering assignment. The gradients flow through a proxy loss that doesn’t directly correspond to the K-Means objective, creating a mismatch between training and inference. The CID module claims to use “counterfactual inference” but implements a simple ablation study. This is not genuine counterfactual inference in the causal inference sense. The work lacks comparison with attention-based MIL with k-means preprocessing and existing disentanglement methods adapted for MIL. Two of four datasets are private and cannot be verified. Results on these datasets show suspiciously large gaps, raising concerns about overfitting. The paper claims existing methods lack structural causal models to build disentangled relationships, yet recent work explicitly uses SCMs: Causal multiple instance learning, Interventional bag multi-instance learning.
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 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?

Papers #366 and #426 use shared Anonymous Repository Link
Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2026/en/REVIEWER-GUIDELINES.html

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

(4) Weak Accept — marginally above the acceptance threshold, but would not mind if rejected, dependent on rebuttal
Please justify your recommendation. What were the major factors that led you to your overall score for this paper?

This paper combines standard metric learning, K-Means clustering, and simple ablation-based counterfactual analysis without genuine methodological innovation. The core technical claim (PSD-LFG) is mathematically flawed and not novel. The counterfactual inference is superficial ablation testing mislabeled as causal reasoning. Experimental results on private datasets cannot be verified, and improvements on public datasets may reflect suboptimal baselines rather than genuine advances. The interpretability claims lack quantitative validation. Moreover, the supplementary materials and the shared shared code may break the conference rules.
Reviewer confidence

Confident but not absolutely certain (3)
[Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.

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

The response resolve my concerns. I support acceptance.

Review #3

Please describe the contribution of the paper

This paper proposes a new WSI classification framework called PG-CIDL. The method is designed to reduce three types of entanglement in MIL based pathology models, spatial, semantic, and decision entanglement. It does this through three stages: latent factor grouping, cluster reasoning based disentangling, and instance effect re weighting. The overall contribution is a more interpretable and better performing WSI representation framework.
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.

This paper proposes a new WSI classification framework called PG-CIDL. The method is designed to reduce three types of entanglement in MIL based pathology models, spatial, semantic, and decision entanglement. It does this through three stages: latent factor grouping, cluster reasoning based disentangling, and instance effect re weighting. The overall contribution is a more interpretable and better performing WSI representation framework.
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 causal interpretation is not fully convincing. Main concern is that the causal language feels stronger than the evidence provided. The leave one group out masking strategy is useful, but it does not fully establish causal validity in a real clinical sense. Also not fully convinced by the fixed assumption that each WSI can always be decomposed into exactly three groups: tumor, microenvironment, and background. That design choice may be too restrictive. In addition, some important details are pushed to the supplementary material, which makes it harder to fully judge the method from the main paper alone. Reproducibility is also somewhat limited since the code is not yet available.
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 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.

(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 this paper a weak accept because I think it addresses an important problem and presents a technically interesting method with strong empirical results. The method appears thoughtful, and the experiments are broader than average, with multiple datasets, strong baselines, ablations, and qualitative analysis. At the same time, I do have some hesitation about the strength of the causal framing, the fixed three group assumption, and the limited reproducibility details in the main paper. So overall, I see it as a solid paper that is above the threshold, though not without weaknesses.
Reviewer confidence

Confident but not absolutely certain (3)
[Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.

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

The authors have comprehensively addressed all of my concerns. I would like to thank them for their detailed responses.

Author Feedback

For Meta Reviewer: Please kindly refer to R1.2, 1.3, R2.1, 2.2 and R3.1 for your concerns.

Reviewer #1: R1.1 The main paper and supplementary one are to comply with the MICCAI dual-submission policy for similar topics (see official guidelines). Critical information, datasets(section4) and CID idea(section3) are available via the abstract link.

R1.2 Instances may be mixed in the original Euclidean space. Thus, we apply the positive semi-definite transformation via W=(A^T)A to project features into a new metric space ||A(xi-xj)||_2 for measurement. Besides the reported inter-group separation term d_reg, we also use an intra-cluster compactness term d_compact, which minimizes the average distance between instances and their assigned cluster centroids. (See abstract link for details).

R1.3 The choice K=3 is a coarse pathological prior rather than exactly three biological subtypes. In SCC setting, these three groups correspond to tumor, microenvironment or background such as blur, bubbles or staining artifacts. In other tasks, they can be coarsely divided into lesion, non-lesion and background. This partition is common across WSI analysis and generalizes across datasets.

R1.4 Ablation experiments on CAMELYON16 shows that, compared with the baseline, PG-CIDL without PSD-LFG improves ACC/AUC by +2.86/+3.83, PG-CIDL without CID improves ACC/AUC by +5.71/+2.77, and the full PG-CIDL improves ACC/AUC by +8.57/+6.94.These results confirm that both PSD-LFG and CID also contribute on a public benchmark.

Reviewer #2: R2.1 Besides d_reg, we omitted the intra-cluster term d_compact, which minimizes the average distance between instances and cluster centroids, completing the K-means objective (See abstract link: section2 for details). Also, background may has category-related meaning and thus contribute to prediction (e.g., vessel and lymphatic reaction).

R2.2 We clarified our paradigm and codes of counterfactual inference: min{L_ce(f(Z_final), y’) + d(Z_final, Z)} in abstract link: section2.Here the minimal optimization generates the smallest perturbed counterfactual representations Z_final while successfully making a counterfactual prediction f(Z_final) to a counterfactual label transition y’. Specifically, y’ is for the pseudo labels from the second argmax of original logits f(Z).

R2.3 PSD-constrained grouping and CID can be plug-and-play modules for commonly used MILs. On CAMELYON16, AUC improves by +3.68 for ABMIL and +6.37 for TransMIL. On AMU-LSCC, ABMIL improves by +3.62 ACC / +5.96 AUC. On AMU-CSCC, ABMIL gains +1.89 ACC, and TransMIL gains +0.48 AUC. These show our methods are not tied to one backbone.

R2.4 Two private datasets had been released and will be verified when camera-ready. On public CAMELYON16, PG-CIDL achieves 0.8857 ACC / 0.9282 AUC, improving over the strongest baseline by +5.71 ACC / +10.29 AUC. Same for DHMC-LUNG. All methods use the same Swin-T features and evaluation protocol. Thus, the gains cannot be attributed only to private-data overfitting.

R2.5 Existing works such as IBMIL and MFC-MIL do introduce SCMs ideas. Our aim is that PG-CIDL explicitly combines: (i) PSD latent factor grouping, (ii) group-level intervention-style effect estimation for semantic disentanglement, and (iii) effect-based decision re-weighting to address spatial, semantic, and decision entanglement within one WSI MIL framework.

Reviewer #3: R3.1 PG-CIDL does not claim clinical causal validity but still has useful interpretability since it evaluates how much the prediction changes when each learned group is removed. Regarding K=3, we use it as a coarse pathological prior rather than a rigorous taxonomy. The model separates diagnostic tissue, microenvironment/non-lesion tissue, and background/artifacts.

R3.2 We will move the CID idea, dataset statistics and training details to the main/camera-ready version where space permits. We have also separated the anonymous code material for PG-CIDL to improve reproducibility.

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 three‑phase disentangled representation learning framework for WSI‑based MIL to reduce spatial, semantic, and decision entanglement. The reviewers agree that the three types of entanglement (spatial, semantic, decision) in existing MIL frameworks are valid and clinically meaningful challenges and recognize the strong empirical results. However, the reviewers also raised concerns about the oversimplified and rigid clustering assumption with a fixed number of clusters (K=3), the unconvincing causal / counterfactual claim, and the reproducibility and reliance on private datasets. Please respond to the main concerns of the reviewers, including the following ones: How does CID handle clustering errors, why is K=3 a sufficient or justified decomposition for all WSIs, how the proxy loss relates to the actual clustering objective and whether the gradient flow correctly optimizes the intended grouping, what is the genuine counterfactual/causal claim in CID, and how does it differ from a simple ablation.
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 rebuttal has addressed most concerns of the reviewers and all three reviewers raised the score from weak accept to accept.

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.

After considering the reviews and rebuttal, I recommend acceptance. The authors have adequately addressed the main concerns regarding the grouping strategy, intervention module, and methodological clarity. The rebuttal provides additional technical explanations, ablation studies, and cross-backbone evaluations showing that both proposed components contribute to performance gains.

While some aspects, such as the justification of the group number and the discussion of related causal MIL methods, could be better explained in the final version, these are primarily issues of presentation rather than fundamental weaknesses. The method is technically sound, demonstrates consistent improvements on both public and private datasets, and offers a useful combination of semantic grouping and intervention-based interpretability for WSI analysis.

Overall, the contribution is meaningful and sufficiently validated to merit 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.

The paper presents a well-motivated framework to address multiple forms of entanglement in MIL-based WSI representation learning, including spatial, semantic, and decision entanglement. Reviewers appreciated the overall framework design, the integration of PSD-based metric learning and disentanglement strategies, and the strong experimental performance across multiple datasets. Overall, despite some remaining limitations regarding the causal framing and fixed clustering assumption, the paper presents a technically solid and practically meaningful contribution for interpretable WSI representation learning, supported by comprehensive experiments and improved performance. I support acceptance.

back to top

Positive Semi-definite Group-Aware Instance Disentangled Learning for WSI Representation

Author(s):