List of Papers Browse by Subject Areas Author List
Abstract
Existing weakly supervised semantic segmentation (WSSS) methods in computational pathology commonly use a multi-stage workflow involving class activation map (CAM) generation, offline pseudo-mask refinement, and fully supervised retraining. While effective, this decoupled workflow can incur high computational training costs and is vulnerable to error propagation, where local texture biases in shallow CNN layers may generate false-positive artifacts that subsequent refinement steps fail to correct. To mitigate these issues, we propose the Single-Stage Hierarchical Rectification (SSHR) framework. Rather than relying on post-hoc CAM refinement, SSHR rectifies intermediate feature representations during the forward pass. We introduce a Hierarchical Feature Rectification Module (HFRM) that uses deep global semantic context to guide CAM-producing shallow and intermediate features, suppressing local anomalies before activation maps are formed. This mechanism produces refined activation maps directly within a single training loop. Experiments on the LUAD-HistoSeg and BCSS datasets demonstrate that SSHR outperforms state-of-the-art multi-stage methods. Furthermore, SSHR reduces training duration by roughly 2 to 5 times. This efficiency lowers computational overhead and supports faster model development for large-scale histopathology workflows. The code is available at: https://github.com/trongduc-nguyen/SSHR.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/3170_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to the Code Repository
https://github.com/trongduc-nguyen/SSHR
Link to the Dataset(s)
LUAD-HistoSeg: https://drive.google.com/file/d/1lWAeCp6UN30VRVmqv97kA2sJ1Pp2frhC/view
BCSS-WSSS: https://drive.google.com/file/d/178eSM9xs5jITt5P2kjaswDlJzwlU5gps/view
BibTex
@InProceedings{NguTro_SingleStage_MICCAI2026,
author = { Nguyen, Trong Duc AND Nguyen, Hoang Long AND Do, Thanh Ha AND Pham, Huy Hieu},
title = { { Single-Stage Hierarchical Rectification for Weakly Supervised Histopathology Segmentation } },
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 proposes a single-stage feature rectification framework for weakly supervised semantic segmentation, aiming to address the error propagation issue in conventional multi-stage CAM-based pipelines. The method introduces semantic guidance and spatial refinement modules to improve feature quality and segmentation consistency within a unified 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.
The paper presents a well-motivated problem formulation, identifying error propagation in multi-stage CAM-based pipelines as a key limitation. The proposed single-stage framework provides a coherent solution aligned with the stated problem. It demonstrates strong empirical performance across multiple benchmarks, showing consistent improvements over existing CAM-based approaches. The diagnostic analysis (e. g. , backbone scaling instability) offers useful insights into limitations of current methods and strengthens the motivation of the work.
- 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 does not clearly discuss the provenance of its modules. While each component is described in terms of functionality, it remains unclear whether these designs are newly proposed or adapted from prior work such as CAM refinement methods (e. g. , IRNet, SEAM) or spatial regularization approaches (e. g. , CRF-based methods). The design choices are not sufficiently justified. The paper explains how each module works, but does not explain why these specific formulations are selected over alternative approaches such as attention-based refinement, affinity learning, or contrastive methods. The problem formulation may be somewhat over-generalized. The paper attributes instability to multi-stage pipelines, but the provided evidence focuses mainly on CAM-based methods. It is unclear whether this issue applies to multi-stage designs in general or is specific to CAM-based supervision. The evaluation is limited to CAM-based methods. Including non-CAM weakly supervised approaches would provide a more comprehensive comparison.
- 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 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.
(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 novelty is somewhat limited at the methodological level. The proposed modules largely build upon existing concepts such as channel attention, global pooling, and large-kernel convolution, without introducing fundamentally new mechanisms. Additionally, the evaluation focuses mainly on performance gains, with limited analysis of failure cases or robustness across diverse scenarios. In summary, the work is solid, well-executed, and practically useful, but its contribution is more incremental in nature, which led to the given score.
- 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.
The rebuttal addresses my concern about module provenance reasonably well by clarifying that HFRM rectifies CAM-producing features before CAM generation, rather than adapting IRNet/SEAM/CRF-style post-hoc refinement. The explanation of GSR and CH is also helpful. However, the justification of the specific design choices remains somewhat qualitative. The efficiency argument for avoiding attention/affinity-based alternatives is reasonable, but no direct evidence is provided. I also appreciate that the authors acknowledge the scope limitation, since the evidence mainly supports CAM-based histopathology WSSS rather than multi-stage weakly supervised segmentation in general. The lack of non-CAM weakly supervised comparisons also remains a limitation. Overall, I think the response partially addresses my concerns. I would keep a positive recommendation, while encouraging the authors to narrow the claims and explicitly discuss the remaining limitations.
Review #2
- Please describe the contribution of the paper
This paper propose a single-stage hierarchical rectification (SSHR) framework for weakly supervised histopathological image segmentation. SSHR outperforms many multi-stage advancements on both the LUAD-HistoSeg and BCSS benchmarks. Also, SSHR reduces total training duration by 2~5 times compared to multi-stage pipelines.
- 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.Single-stage WSSS framework in histopathological scenario is interesting. 2.HFRM purifies CAMs directly during the forward pass. 3.It greatly reduces training time, which is essential for clinical deployment.
- 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.ResNet38 is served as the classification network, it has many layers. Which one provides the shallow features F_low? Also, why is that specific layer chosen? 2.The paper states that W_1 and W_2 in Eq. 1 are learnable, but it doesn’t explain how they are computed. Also, the dimensions don’t match: W_1 is D×C/8 while W_2 is C×C/8, making multiplication impossible. 3.GSR works by adjusting channel-level activations. But channel attention is already well studied. For example, AMR [1] used Gaussian functions to reactivate weak regions, ARML [2] used exponential functions to reshape CAMs, and MLPS [3] used progressive dropout to focus on low-activation areas. So why does GSR, which only uses simple MLPs, claim to capture global semantics? A clear analysis with these existing methods is needed. [1] Activation modulation and recalibration scheme for weakly supervised semantic segmentation, AAAI, 2022.[2] Mining gold from the sand: weakly supervised histological tissue segmentation with activation relocalization and mutual learning, MICCAI, 2024.[3] Multi-layer pseudo-supervision for histopathology tissue semantic segmentation using patch-level classification labels, MedIA, 2022.4.The same problem recurs in Eq. 4: how are gamma_s and gamma_c defined? 5.It is commendable that Table 3 examines the performance drop of different multi-stage SOTA methods. However, I want to know why is the single-stage performance of SSHR under different backbones not reported? I think this is also an important part of evaluating model robustness.
- 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?
My recommendation is based on the section of weaknesses. If the authors can address these weaknesses in their rebuttal, I will increase my score.
- 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.
The authors addressed my concerns. Therefore I recommend to accept. I hope authors to revise their paper based on my suggestions in their final version.
Review #3
- Please describe the contribution of the paper
1) The paper proposes a single-stage weakly supervised histopathology segmentation framework, SSHR, to address two practical challenges. 2) It introduces a Hierarchical Feature Rectification Module (HFRM), which uses the global semantics of deep features to generate channel weights, thereby suppressing noisy shallow-feature responses that are inconsistent with global semantics. In addition, a large-kernel depthwise convolution is used to enforce spatial homogenization, allowing isolated local noise to be assimilated by the surrounding dominant tissue patterns.
- 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 problem addressed by the paper is of practical significance. 2) The proposed method is relatively simple, yet the results indicate that it achieves reasonable accuracy with high computational efficiency. 3) The effectiveness of the model was demonstrated on two datasets. 4) The paper provides source code, which supports good reproducibility.
- 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 paper shows only limited novelty. Its contribution is better characterized as a targeted combination and forward integration of existing modules for histopathology WSSS, rather than a fundamentally new method with strong conceptual innovation. 2) The theoretical analysis of the method is relatively weak. For example, it remains unclear why the channel suppression mechanism would not mistakenly suppress truly relevant but rare small regions, or why the large-kernel smoothing operation would not systematically damage boundary details. 3) The method carries potential risks of over-smoothing and suppression of small targets. 4) Comparisons with other single-stage methods are limited.
- 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.
(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 proposed method shows limited conceptual novelty and is better viewed as an integration of existing ideas. Moreover, the theoretical and empirical support is insufficient to rule out over-suppression, boundary over-smoothing, and to establish competitiveness against prior single-stage methods.
- 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 authors partially clarify the positioning of SSHR and its risk-mitigation design. However, they still provide insufficient evidence regarding the boundaries of novelty, theoretical soundness, validation of small-object and boundary-related risks, and fair single-stage comparisons.
Author Feedback
We sincerely thank the reviewers for recognizing SSHR’s practical value, efficiency, and strong empirical performance. We address the main concerns below.
For R1, HFRM is not adapted from IRNet, SEAM, or CRF-style refinement, which mainly correct pseudo-labels or predictions after CAMs are generated. SSHR instead rectifies CAM-producing features during the forward pass, before local texture artifacts become pseudo-label errors. This design targets two CAM-formation failures in histopathology. First, low/intermediate features preserve tissue details but can activate on locally discriminative textures without sufficient global context; GSR uses the deepest classification-supervised feature as a patch-level semantic prior to condition these CAM-producing channels before CAM generation. Second, even semantically valid classes may produce isolated spatial outliers; CH adds a lightweight local-contiguity prior through depthwise context filtering. We choose these lightweight operations because SSHR targets efficient rectification for histopathology WSSS; attention or affinity learning is possible, but dense pairwise modeling can substantially increase cost. We agree with R1’s valuable scope suggestion. Our evidence focuses on CAM-based histopathology WSSS, where CAM pipelines remain the dominant directly comparable setting due to practical efficiency. We will discuss broader non-CAM comparisons as important future work.
We thank R2 for raising these thoughtful points and apologize for any ambiguity. F_low denotes the CAM-producing intermediate features. In ResNet38, HFRM is applied to the 56x56, 28x28-1, and 28x28-2 features with 256, 512, and 1024 channels. These multi-scale features are selected because they capture complementary tissue details and each has its own CAM head contributing to the multi-scale classification loss, thus receiving direct image-level supervision. In Eq.1, v_g is in R^D, W1 is in R^{D/8 x D}, and W2 is in R^{C x D/8}. In Eq.4, gamma_s and gamma_c are zero-initialized learnable scalar residual coefficients. Regarding AMR/ARML/MLPS, these works improve CAM completeness or pseudo-supervision through weak-region mining mechanisms such as compensation CAMs, activation relocalization, and progressive dropout/multi-layer pseudo-supervision. Rather than claiming the MLP itself captures global semantics, GSR uses the deepest classification-supervised context and projects it into scale-specific channel weights for lightweight feature rectification. Table 3 diagnoses Stage-2 retraining instability, so SSHR is outside this diagnostic table. We agree that evaluating robustness by changing the Stage-1 backbone is meaningful. In this study, we keep the common ResNet38 Stage-1 backbone used by most prior methods to ensure a fair comparison and isolate the effect of HFRM rather than backbone capacity. Thank you again for these helpful comments; we will clarify the notation and global-context wording in the revision.
For R3, SSHR’s main novelty lies in showing a promising single-stage direction for histopathology WSSS, where forward-time feature rectification can achieve strong performance without relying on post-hoc refinement. We sincerely thank R3 for highlighting the important risks, which directly motivate our residual rectification and zero-initialized scaling design. GSR performs soft channel reweighting rather than hard spatial removal, and gamma_s/gamma_c allow the model to start from the original backbone. The ablation confirms that excessive smoothing is harmful, highlighting the importance of an appropriate kernel size. Regarding single-stage comparisons, pathology-specific single-stage WSSS without post-processing remains underexplored. We therefore compare SSHR against stronger multi-stage SOTA, whose Stage 1 already performs comparable CAM generation and whose final results further benefit from Stage-2 retraining/backbones. Despite using only one stage, SSHR still outperforms their final results.
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.
This work presents a single-stage hierarchical rectification framework for weakly supervised histopathological image segmentation. Initially, the reviews received are mixed, with reviewers acknowledging the interesting scenario presented (and of practical significance), and a well-motivated idea, which is supported by strong empirical evidence. Nevertheless, they have identified important limitations that hinder the impact of the proposed approach. In particular, the very limited novelty, the lack of a clear positioning/contribution (e.g., #R1: why the different modules are selected and why they should work?), absence of relevant comparisons to support some of the claims (e.g., claim to capture global semantics: compare to [1-3]), as well as the lack of comparisons to non-CAM approaches, undermine the overall contribution/impact of this work. I strongly encourage the authors to respond to these concerns in a rebuttal.
- 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.
Overall, authors have addressed several of the most important concerns raised during the rebuttal period. While some limitations remain (e.g., I do agree that authors should have included non-CAM weakly-supervised methods in their empirical validation, or better justification of the specific design choices), I believe this work has merits to be accepted. Therefore, I recommend Accept, and strongly encourage authors to integrate their responses to reviewers concerns in the camera ready.
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 reviewers generally agree on the practical value and solid performance of the proposed method. While there are still questions about the insufficient validation, most concerns were addressed in the rebuttal.
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.
The paper received mixed reviews during the first round of evaluation and it was invited for rebuttal. The authors submitted a rebuttal to address the raised concerns, clarifying different points raised by the reviewers. All the reviewers appreciated the answers with two of them keeping their positive scores and one maintaining ther negative. However, two reviewers indicated that their concerns had been only partially addressed, leaving some open questions about the paper, including better justification of the specific design choices, novelty, and comparisons. The AC read the paper, reviews, rebuttal, and final justification, and agrees that not all the reviewers’ concerns had been properly addressed, making the paper not yet ready for publication.