List of Papers Browse by Subject Areas Author List
Abstract
Histopathological image segmentation is critical for computer-aided pathological diagnosis and prognosis analysis. Although weakly supervised methods reduce the reliance on pixel-level annotations, they are often limited by low-quality pseudo-masks and accumulated label noise. In particular, existing Class Activation Mapping (CAM) approaches are vulnerable to spurious correlations, and conventional networks struggle to jointly preserve fine-grained tissue morphology and global semantic consistency. To address these issues, we propose C\textsuperscript{2}RM-Seg, a unified framework that tightly integrates causal inference with foundation model priors. In the classification stage for pseudo-label generation, we introduce a Causal Counterfactual Reasoning Module (C\textsuperscript{2}RM), which projects features into latent causal factors and learns a causal structure matrix to enable counterfactual interventions. By attenuating confounding contextual effects, C\textsuperscript{2}RM refines CAM activations and yi
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/4147_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{ZhaHua_C²RMSeg_MICCAI2026,
author = { Zhang, Hualong AND Feng, Siyang AND Huan, Zihan AND Qian, Yi AND Liu, Zhenbing AND Lan, Rushi AND Pan, Xipeng},
title = { { C²RM-Seg: Causal Counterfactual Reasoning with Structural-Semantic Priors for Weakly Supervised Histopathological Tissue Segmentation } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 16880},
month = {September},
page = {pending}
}
Reviews
Review #1
- Please describe the contribution of the paper
The paper attempts to mitigate pseudo-label noise by applying a Structural Causal Model and counterfactual intervention to deconfound spurious correlations in histopathology images. This dual-path architecture combines traditional ResNeSt structural features with semantic priors from DINOv3 using a gating mechanism to balance global semantics and morphological details. To manage the inherent noise in weak supervision, a margin loss guided by prediction uncertainty is introduced to dynamically balance margin enforcement and confidence penalties.
- 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 leverages a Structural Causal Model to address the long-standing issue of spurious correlations in CAM, providing a more theoretically grounded approach to pseudo-label refinement. The paper capitalizes on the robust representation power of DINOv3 while maintaining fine-grained morphological details through a dual-path integration.
- 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 proposes a Causal Counterfactual Reasoning Module that relies on vague jargon and a simplistic subtraction formula Z^{cf}=Z-AZ without providing any formal mathematical derivation or empirical evidence of its causal decoupling effect. The paper describes a dual-path design that is a straightforward assembly of existing components including DINOv3 and ResNeSt, where the cross-path gating is a standard attention variant that lacks technical innovation.
- 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 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.
(2) Reject — should be rejected, independent of rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
The paper presents a Causal Counterfactual Reasoning Module with an insufficient theoretical foundation, as the core subtraction formula lacks a formal mathematical derivation to justify its effectiveness in deconfounding latent features. The architectural design consists of a limited combination of existing DINOv3 and ResNeSt components, showing a lack of original technical innovation or specific optimization for histopathological structures. The experimental evaluation is incomplete, failing to provide quantitative comparisons with mainstream weakly supervised segmentation methods.
- Reviewer confidence
Very confident (4)
- [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.
Although the rebuttal clarifies several points, the core concerns regarding the insufficient theoretical grounding of the causal formulation and the limited methodological novelty remain unresolved. Therefore, I maintain my reject recommendation.
Review #2
- Please describe the contribution of the paper
This paper propose C2RM-Seg, a two stage framework for weakly supervised histopathological tissue segmentation. The authors introduce C2RM to suppress confounding context and get better CAM outputs. Then use a dual-path structural-semantic architecture to handle both boundary and semantics. Finally, they use an uncertainty aware UGM loss to handle the pseudo label noise. The paper demonstrate outstanding performance on BCSS-WSSS and LUAD-HistoSeg datasets.
- 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 is clearly written and easy to follow, with a well-structured presentation. 2.The proposed C2RM offers a novel way to refine features for generating improved CAMs. 3.The UGM loss is a useful design for handling the uncertainty of pseudo-masks derived from CAMs
- 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.Regarding the C2RM part, the paper assumes that the projected feature groups correspond to meaningful latent factors, but no evidence is provided to show that these factors indeed capture distinct and meaningful tissues or patterns. It is also unclear what prevents these factors from being redundant or collapsing. Furthermore, in Eq.4, the method subtracts the co-varying component, why this component is regarded as noise rather than useful foreground?
2.Regarding the segmentation part, the choice of ResNeSt and DINOv3 is quite questionable, given that there already exist many strong pathology encoders such as UNI and CONCH. Also, the rationale that DINOv3 captures semantic and ResNeSt captures structure remains a claim without any evidence to support.
- 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?
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 purpose of each component is clear, and together they form a reasonable framework for improving segmentation performance. However, some of the design rationales still appear to rely more on empirical intuition than on direct supporting evidence. Overall, I would give this paper a weak accept.
- 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 concerns are basically solved.
Review #3
- Please describe the contribution of the paper
1.The paper proposes a Causal Counterfactual Reasoning Module for CAM debiasing, aiming to suppress confounding background signals and generate high-fidelity pseudo-labels. 2.It further introduces a Dual-Path Structural–Semantic Architecture that integrates ResNeSt with a frozen DINOv3 backbone to preserve boundary details while incorporating global semantic information, together with an uncertainty-gated margin (UGM) loss for noise-robust optimization. 3.The results were validated on two datasets.
- 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 motivation for addressing this problem is well grounded and of practical value. 2.The overall methodological pipeline is complete, as it first improves the source of pseudo-labels, then enhances the segmentation representation, and finally mitigates residual noise. 3.The experimental results and visualizations are relatively convincing. 4.The paper is logically well structured, with a clear and coherent organization.
- 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 learn a trainable relationship matrix in the latent factor space and then perform debiasing through linear subtraction. This subtraction appears to be only a linear correction, rather than a truly verifiable counterfactual intervention. 2.The counterfactual correction formula is overly simplified, whereas confounding in real histopathological images is often much more complex. 3.Matrix A appears to play an important role in the model, but its analysis in the results section is rather limited. 4.The experiments support performance improvement, but they do not fully substantiate the claimed mechanism. The core claim of the paper is that it removes staining-driven confounding; therefore, the most compelling evidence would be experiments showing that causal deconfounding itself is the key factor behind the improvement. 5.The ablation studies and sensitivity analyses are still not sufficiently comprehensive.
- 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.
(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 practical significance of the problem addressed, the completeness of the method, and the clarity of the overall presentation.
- 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 provide useful additional evidence, including an analysis of the learned relationship matrix, a staining robustness stress test, and sensitivity results for the number of latent factors.
Author Feedback
We sincerely thank the reviewers for their insightful feedback. 1.Mathematical Foundation and Empirical Evidence of Z^{cf} = Z - AZ (R1,R2,R3): This is not a heuristic, but strictly derived from Structural Causal Models (SCMs). In SCM, features are modeled as Z = AZ + \epsilon, where AZ quantifies confounding effects (e.g.,staining bias) via dependency matrix A, and the exogenous variable $\epsilon$ represents the pure, unconfounded causal signal (e.g.,intrinsic morphology). Performing counterfactual intervention to block causal paths from confounders mathematically equates to isolating \epsilon. By simple rearrangement, \epsilon = Z - AZ. Thus, Z^{cf} = Z - AZ is the exact algebraic implementation of extracting deconfounded features by stripping away learned spurious correlations. While this establishes the theoretical soundness, the empirical evidence confirming that AZ genuinely captures confounding noise rather than useful foreground features in practice is detailed in Point 3 (Weight Analysis of Matrix A) and Point 4 (Staining Stress Test) below. To address R1’s concern regarding evaluation, we emphasize that the methods compared in our experiments are highly representative and top-performing mainstream in WSSS. 2.Latent Factor Meaning & Anti-Collapse (R2): Multi-class CAM optimization inherently forces channel groups to learn distinct patterns. Although previously omitted due to space limits, our analysis shows C2RM reduces inter-factor cosine similarity from 0.68 to 0.21.Alongside t-SNE (Fig. 3), this proves decoupling without collapse. 3.Why Subtract Noise, Not Foreground? (R2, R3): Global staining acts as a network “shortcut”, whereas foreground morphology is localized. Although omitted previously due to space limits, our analysis of matrix A shows its high-weight connections are sparse and strictly directed from background to foreground factors. Thus, subtracting AZ strips away spurious shortcuts rather than the foreground. 4.Proof of Causal Deconfounding (R1, R3): Although omitted previously due to space limits, we conducted a staining robustness stress test (extreme color jittering and Macenko normalization) to prove our mechanism. Under severe confounding, the baseline mIoU collapsed by 5.24%, whereas C2RM-Seg showed a mere 1.57% drop. This verifies that causal deconfounding is the core driver of our performance gains. 5.Architectural Novelty & Cross-Path Gating (R1): Our dual-branch design is not mere concatenation. The major bottleneck in WSSS for pathology is pseudo-label boundary. Our Cross-path Gating addresses this by using spatial confidences derived from high-frequency local structures (ResNeSt) to dynamically “throttle” the injection of DINOv3’s global semantics. This prevents smooth global features from blurring fine-grained tissue boundaries. 6.Why DINOv3 Instead of UNI or CONCH? (R2): We conducted comparative experiments using UNI and CONCH as the semantic branch. Under identical settings, UNI and CONCH achieved mIoUs of 78.41% and 78.75%, respectively, both lower than DINOv3 (79.62%). While UNI/CONCH excel at image-level classification, DINOv3’s self-supervised objectives (mask prediction) make it far superior for pixel-level dense prediction. 7.Proof of “Structure vs. Semantic” Roles (R2): Though omitted earlier due to space limits, our Fourier spectrum and bIoU analyses on single branches show ResNeSt features are rich in high frequencies. Conversely, DINOv3 exhibits strong low-frequency semantics but blurry boundaries (bIoU drops by 4.2%). Their complementarity under our gating mechanism validates our design. 8.Comprehensive Ablations & Sensitivity (R3): Omitted from the original submission due to strict space constraints, we conducted sensitivity analyses for the number of latent factors K in {4, 8, 16, 32}. The model achieves optimal performance at K=8 (mIoU 79.62%). Capacity is limited at K=4 (77.2%), but performance remains highly stable at K=16, 32 (79.12%).
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.
Please address the concerns about Weak theoretical grounding of the counterfactual subtraction, Insufficient evidence that the mechanism works as claimed.
- 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.
two accept and one reject.
i read through and would accept subject to minor
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.
C2RM-Seg addresses an important weakly supervised histopathology segmentation problem and presents a complete pipeline: CAM refinement, structure–semantic segmentation, and uncertainty-aware training. The submitted experiments compare against relevant WSSS and foundation-model baselines on two public datasets, report variance, boundary metrics, efficiency, and include meaningful ablations. While R1 raises valid concerns that the causal interpretation of Z^{cf} = Z − AZ is stronger than what is rigorously proven, the rebuttal clarifies the intended SCM-inspired approximation, and the original manuscript already provides component ablations and feature visualization supporting its practical effect. I do not consider the additional rebuttal-only stress tests or encoder comparisons, but the submitted evidence is still sufficient. The method should revise claims to “SCM-inspired/counterfactual feature correction” rather than fully verified causal deconfounding, and better explain the assumptions behind the latent factor graph.
Meta-review #3
- 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.
I recommend rejection. While the paper addresses an important problem in weakly supervised histopathological tissue segmentation, the main technical claims are not sufficiently convincing. In particular, the proposed causal counterfactual reasoning module is presented as causal deconfounding, but the learned factor matrix and subtraction-based intervention do not provide a rigorous or empirically validated causal model. As a result, the central motivation of removing staining-driven confounders remains largely asserted rather than demonstrated.
The method also appears to combine several existing ideas—CAM refinement, foundation-model features, dual-branch fusion, gating, and uncertainty-aware loss—without clearly establishing which component is truly novel or necessary. The experimental validation is limited to two datasets, and the improvements over strong recent WSSS baselines are relatively modest given the added complexity. More importantly, the paper lacks targeted evidence on stain/domain robustness, pseudo-label quality, causal validity, and generalization across centers or scanners, which are essential for supporting its main claims.
Overall, although the direction is relevant, the current submission overstates its causal contribution and does not provide sufficiently strong methodological or experimental evidence to justify acceptance.