List of Papers Browse by Subject Areas Author List
Abstract
Histopathology image analysis is critical yet challenged by the demand of segmenting tissue regions and nuclei instances for tumor microenvironment and cellular morphology analysis. Existing studies focused on tissue semantic segmentation or nuclei instance segmentation separately, but ignored the inherent relationship between these two tasks, resulting in insufficient histopathology understanding. To address this issue, we propose a Co-Seg framework for collaborative tissue and nuclei segmentation. Specifically, we introduce a novel co-segmentation paradigm, allowing tissue and nuclei segmentation tasks to mutually enhance each other. To this end, we first devise a region-aware prompt encoder (RP-Encoder) to provide high-quality semantic and instance region prompts as prior constraints. Moreover, we design a mutual prompt mask decoder (MP-Decoder) that leverages cross-guidance to strengthen the contextual consistency of both tasks, collaboratively computing semantic and instance segmentation masks. Extensive experiments on the PUMA dataset demonstrate that the proposed Co-Seg surpasses state-of-the-arts in the semantic, instance and panoptic segmentation of tumor tissues and nuclei instances. The source code is available at https://github.com/xq141839/Co-Seg.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2025/paper/0931_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to the Code Repository
https://github.com/xq141839/Co-Seg
Link to the Dataset(s)
N/A
BibTex
@InProceedings{XuQin_CoSeg_MICCAI2025,
author = { Xu, Qing and Duan, Wenting and Chen, Zhen},
title = { { Co-Seg: Mutual Prompt-Guided Collaborative Learning for Tissue and Nuclei Segmentation } },
booktitle = {proceedings of Medical Image Computing and Computer Assisted Intervention -- MICCAI 2025},
year = {2025},
publisher = {Springer Nature Switzerland},
volume = {LNCS 15960},
month = {September},
page = {133 -- 143}
}
Reviews
Review #1
- Please describe the contribution of the paper
(1) This paper proposes a Co-Segmentation Framework (Co-Seg) for collaborative tissue and nuclei segmentation in histopathology. Co-Seg allows multi-segmentation tasks to mutually enhance each other. (2) Co-Seg introduces a RP-Encoder to capture task-specific priors. Furthermore, MP-Decoder is designed to enable cross-task guidance through bidirectional feature interactions. (3) Co-Seg’s collaborative segmentation in tissue and nuclei segmentation demonstrates mutual improvement.
- 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.
Co-Seg proposes a closed-loop bidirectional interaction paradigm for tissue and nuclei segmentation, enabling mutual enhancement between semantic and instance tasks. Co-Seg’s collaborative segmentation approach leverages feature priors from different tasks to enhance understanding and performance in individual segmentation tasks.
- 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’s claims of task complementarity lack theoretical proof of convergence and fail to address potential task conflicts, risking performance saturation. Empirical loss weight configurations (e.g., fixed λ₁=2, λ₂=1) lack justification for inter-task balance, while static query sharing in the decoder neglects adaptive mechanisms (e.g., dynamic attention), limiting adaptability in heterogeneous tissue regions. (2) The paper identifies the negative impact of task decoupling but lacks substantial evidence or detailed analysis to support this claim. (3) Data description errors: The article references the PUMA dataset, which was initially reported to contain 103 images, whereas the actual dataset provided 205 correctly labeled images during the competition. After the competition, the dataset was updated to include 206 images, 1024x1024 resolution. The paper’s unclear usage of the PUMA dataset raises concerns about the reproducibility of its results. (4) Insufficient experimental coverage: The paper relies solely on the PUMA dataset for experimentation, raising concerns about its usage due to a lack of clarity. Without an official test set, the authors’ use of self-splitting is acceptable, but the absence of multiple folds and replication experiments (mean+std) undermines the method’s validity and stability. (5) Ablation study is incomplete: The ablation study lacks parallel task evaluations, which is significant as it hinders the interpretation of results and limits the understanding of the collaborative framework’s contributions. (6) Incorrect citations and writing errors: The paper’s figures and legends exhibit inconsistencies, such as the RP-Encoder’s integration of mask logits with image embeddings not being clearly represented in Fig. 2(b). These discrepancies raise concerns about the clarity and accuracy of the framework’s presentation. Furthermore, the paper’s references may require verification to ensure proper citation practices.
- 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.
- 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/2025/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?
Co-Seg proposes a novel bidirectional collaborative framework for tissue and nuclei segmentation, leveraging task-specific feature priors to enhance performance. However, its impact is limited by unresolved data transparency (e.g., PUMA dataset inconsistencies), insufficient experimental validation (single dataset, no statistical replication), and incomplete ablation studies. While the methodology is innovative, addressing gaps in reproducibility (standardized dataset splits) and theoretical rigor (task decoupling analysis) is critical. Future work should prioritize cross-dataset testing and precise technical documentation to validate claims.
- 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 author’s rebuttal answers my question to some extent, hoping that the previous shortcomings will be corrected in the final revision.
Review #2
- Please describe the contribution of the paper
The paper introduces Co-Seg, a novel framework that enables collaborative learning between tissue semantic segmentation and nuclei instance segmentation in histopathology images. The key contributions include: 1) a co-segmentation paradigm that formalizes the mutual dependency between these tasks through intertwined conditional probabilities; 2) a region-aware prompt encoder that extracts high-quality semantic and instance prompts; 3) a mutual prompt mask decoder that leverages cross-guidance for collaborative segmentation.
- 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 authors should be commended for identifying and addressing the inherent relationship between tissue and nuclei segmentation tasks, which previously were treated as separate problems. This represents a paradigm shift in how we approach histopathology image analysis
-
The proposed architecture with RP-Encoder and MP-Decoder demonstrates thoughtful design choices. I believe, this could benefit in other computer vision/medical image analysis applications.
-
The quantitative performance improvements are significant.
-
- 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 evaluation is performed on a single PUMa dataset. A better and detailed evaluation on multiple datset could strengthen the paper’s effectiveness. Different tissues might exhibit varying degrees of correlation between tissue structures and nuclear morphology, potentially affecting the effectiveness of the collaborative approach.
-
The co-segmentation paradigm assumes a symmetrical interdependence between tasks, as formalized in Equation (1), but this requires deeper examination. In certain pathological conditions, tissue architecture might be more informative than nuclear features (or vice versa). The framework doesn’t account for potential asymmetry in task importance or reliability, which could limit its effectiveness in complex clinical scenarios.
-
It’s unclear why the semantic prompt helps in instance segmentation and vice versa. It is rather intuitive to use a semantic prompt for a semantic decoder and instance prompt for instance decoder. The motivation for the particular architecture design is not well-supported in the introduction and motivation section.
-
How is the second and third term implemented in terms of code?
-
The effectiveness of the mutual prompting mechanism heavily depends on the quality of the prompts generated by each task. When one task produces low-quality prompts (e.g., due to challenging tissue morphology), these errors propagate to the other task, potentially causing cascading failures. How does this paper address that?
-
- 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.
- 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/2025/en/REVIEWER-GUIDELINES.html
Minor issues:
- Typo: Abstract says - “To end this, we first devise a region-aware prompt encoder (RP-Encoder)…”, should be “to this end..”
- None of the papers that the authors compared against are for joint tissue and nuclei segmentation.
- No evaluation is performed on how incorporating prompting techniques could improve the performance of the existing baselines in literature.
- 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 — could be rejected, dependent on rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
Although the contributions are solid, I’m still skeptical about a few weakness points. I’d like to improve my score if my concerns are thoughtfully addressed in the rebuttal.
Good luck to the authors :)
- 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.
The rebuttal has addressed some of my concerns, but still some concerns remain, especially the points raised by Reviewer 1. I believe, those are valid points and should be taken care of before the paper could be recommended for publication.
This was a really tough decision to make :(
Review #3
- Please describe the contribution of the paper
This paper proposes a collaborative learning-based segmentation framework to simultaneously segment tissue regions and nuclei instances. Leveraging the inherent relationship between these two tasks, a co-segmentation approach is designed to enable mutual enhancement. Specifically, a Region-aware Prompt Encoder (RP-Encoder) is introduced to generate region prompts as prior constraints for both tasks. These prompts are then used by the Mutual Prompt Mask Decoder (MP-Decoder), which employs cross-guidance to predict both segmentation masks while ensuring contextual consistency. Experiments conducted on the PUMA dataset demonstrate that the proposed co-segmentation framework achieves state-of-the-art performance on both tissue and nuclei segmentation tasks.
- 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 is well-motivated, leveraging the inherent contextual relationship between tissue and nuclei segmentation tasks. The proposed framework is primarily built upon region prompts and utilizes cross-attention mechanisms to establish cross-task contextual consistency. The approach demonstrates superior performance on both segmentation tasks. Additionally, the ablation study is relatively thorough and effectively explains the performance contribution of each designed component. Overall, the paper is well-written and easy to follow.
- 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.
I only have a few curious questions and suggestions. 1.Instead of using the predicted masks, what would happen if the ground-truth masks were fed into the RP-Encoder to generate prior constraints? That is, can the ground-truth of one task be used to guide the learning of the other? Is this feasible, and how does the performance compare to the current approach? 2.This work can also be viewed as a task consistency approach, since it enforces consistency between two correlated segmentation tasks (tissue and nuclei). It would be beneficial to discuss related works on task consistency to better position this method within the broader research landscape. 3.It would be more convincing to include an additional dataset to demonstrate the generalizability and effectiveness of the proposed method.
- 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.
- 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/2025/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?
This paper is well-motivate, well-written and easy to follow.
- Reviewer confidence
Somewhat confident (2)
- [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
We appreciate all reviewers for their valuable appreciation of our paper regarding “clear novelty”(R1,R3), “well-motivated”(R2), “impressive effectiveness”(R2,R3) and “thoughtful design”(R3). R1Q1:Convergence of Co-Seg. The tissue and nuclei segmentation tasks are inherently complementary rather than conflicting. Our experiments confirmed the rapid convergence of Co-Seg within 50 epochs during training. R1Q2:Evidence of Task-Decoupled Impact. Tab 1-2 and 4 show that task-decoupled methods [2,4,5,8,12,14,17,20,25,26] on independent tasks are inferior to Co-Seg, demonstrating their disadvantages. R1Q3:Dataset Description. Upon verification, we confirm that our experiments use 206 images. We will revise in the final version to avoid ambiguity. R1Q4:Insufficient Experiment. Tab 1–3 report multiple evaluation protocols with clear performance gains and statistical significance, demonstrating the reliability of our results. R1Q5:Incomplete Ablation Study. The 1st row of Tab 4 shows the baseline with parallel task, allowing us to evaluate the incremental impact of our collaborative designs. Our ablation study is also recognized by R2. R1Q6:Figure Clarity and Citation Accuracy. In Fig 2b, the integration of mask logits with image embeddings is clearly performed via cross-attention, as defined in Eq(5). Moreover, we have verified that our citations are correct. R2Q1:GT Masks as Prior Constraints. Using GT masks as prompts are ideal and would likely improve performance as prior constraints are supervised by the GT masks during training. R2Q2:Related Works on Task Consistency. We will further discuss related works on task consistency in extension. R2Q3:Validation on Additional Datasets. We agree that broader validation is important and will pursue multi-dataset evaluation in extension. R3Q1:Limited Dataset Evaluation. Due to the rebuttal policy, we cannot add new experiments, but we will evaluate additional datasets in the extension. In fact, we have conducted comprehensive experiments on the PUMA dataset that contains histopathology images across diverse pathological stages and is evaluated in three different segmentation tasks. The clear performance gains, statistical significance, thorough ablation and qualitative comparisons shown in Tab 1-4 and Fig 3 have demonstrated the effectiveness of Co-Seg. R3Q2:Potential Asymmetry of Tissue and Nuclear Information. The information of both tasks is shared with each other. As defined at Eq (6-7), mutual prompts allow Co-Seg to adapt to varying levels of informativeness between tasks during training. This cross-guidance avoids potential asymmetry. R3Q3:The Motivation of Mutual Prompts. Semantic and instance prompts are highly correlated in histopathology as tissue semantic regions inform the localization of nuclei instances, and nuclear arrangements help distinguish tissue subtypes. Co-Seg aims to achieve cross-guidance between semantic and instance segmentation. As shown in the 3rd and 4th rows of Tab 4, Co-Seg with mutual prompts outperforms that with parallel prompts, proving its effectiveness. R3Q4:Code Implementation. The 2nd and 3rd terms of Eq(2) and Eq(8) are implemented from line 154-195 and 88-134 of the “model.py” and “train.py” files, respectively. Overall, the main Implementation of Co-Seg is included in both files. R3Q5:Impact of Low-Quality Prompts. Co-Seg adopts ground-truth-supervised prior constraints during training Eq(8), helping the RP-Encoder learn to generate reliable prompts. Tab 1–3 show that Co-Seg outperforms non-prompted methods, even when one task’s prompts may be suboptimal. R3Q6:Comparison with Joint Tissue and Nuclei Segmentation. The panoptic segmentation evaluation shown in Tab 3 considers joint semantic and instance segmentation. R3Q7:Evaluation of Prompting Techniques. The efficiency of incorporating prompting has been proven in the 2nd and 3rd rows of Tab 4, and existing studies [17, 25].
R3Q8:Typo. We will correct identified typos in final version.
Meta-Review
Meta-review #1
- Your recommendation
Invite for Rebuttal
- If your recommendation is “Provisional Reject”, then summarize the factors that went into this decision. In case you deviate from the reviewers’ recommendations, explain in detail the reasons why. You do not need to provide a justification for a recommendation of “Provisional Accept” or “Invite for Rebuttal”.
Paper Summar: This paper tackles the challenge of jointly segmenting tissue regions and nuclei instances in histopathology by introducing a Co-Seg framework that enables collaborative learning between semantic and instance tasks. It comprises two key modules: a region-aware prompt encoder that extracts task-specific priors from initial mask predictions, and a mutual prompt mask decoder that uses bidirectional cross-attention to refine both tissue and nuclei segmentation simultaneously.
Key Strengths: The proposed co-segmentation paradigm is novel, leveraging mutual prompts to exploit interdependencies between tissue and nuclei segmentation. This bidirectional interaction leads to significant quantitative gains over SOTA methods on the PUMA dataset. The architecture is thoughtfully designed and well motivated, and the ablation study demonstrates clear contributions of each component to overall performance.
Key Weaknesses: All experiments are confined to a single dataset without cross-dataset evaluation or statistical replication, raising concerns about generalizability and stability. There is no theoretical analysis of convergence or potential task conflicts under fixed loss weights, and the rationale for certain design choices, such as static prompt sharing, remains under-explained. Inconsistent descriptions of the PUMA dataset and minor citation and writing errors further undermine clarity and reproducibility.
Review Summary: Reviewers commend the innovative collaborative learning approach and its strong empirical performance. They agree the framework is well written and the core idea of mutual prompt guidance is compelling. However, they diverge on the sufficiency of validation: some find the ablation study thorough, while others call for additional datasets and replication experiments. Concerns also surface regarding theoretical grounding of the co-segmentation paradigm and deeper analysis of design motivations and failure modes.
Decision Invite to rebuttal –The manuscript presents a promising collaborative segmentation framework but needs further empirical validation and clearer justification of key design elements.
- 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. You may optionally write justifications for ‘accepts’, but are expected to write a justification for ‘rejects’
N/A
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. You may optionally write justifications for ‘accepts’, but are expected to write a justification for ‘rejects’
This paper introduces a compelling approach to joint tissue and nuclei segmentation through a collaborative framework. Despite limitations in data transparency, experimental breadth, and reproducibility, the methodology is innovative and well-motivated, with strong empirical performance and architectural potential beyond the current application.
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. You may optionally write justifications for ‘accepts’, but are expected to write a justification for ‘rejects’
N/A