List of Papers Browse by Subject Areas Author List
Abstract
Foundation models such as Segment Anything Model 3 (SAM3) enable flexible text-guided medical image segmentation, yet their predictions remain highly sensitive to prompt formulation. Even semantically equivalent descriptions can yield inconsistent masks, limiting reliability in clinical and pathology workflows.
We reformulate prompt sensitivity as a group-wise consistency problem. Semantically related prompts are organized into prompt groups sharing the same ground-truth mask, and a prompt group-aware training framework is introduced for robust text-guided nuclei segmentation. The approach combines (i) a quality-guided group regularization that leverages segmentation loss as an implicit ranking signal, and (ii) a logit-level consistency constraint with a stop-gradient strategy to align predictions within each group. The method requires no architectural modification and leaves inference unchanged.
Extensive experiments on multi-dataset nuclei benchmarks show consistent gains under textual prompting and markedly reduced performance variance across prompt quality levels. On six zero-shot cross-dataset tasks, our method improves Dice by an average of 2.16 points. These results demonstrate improved robustness and generalization for vision-language segmentation in computational pathology.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/5066_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to the Code Repository
https://github.com/yonghongwu/sam3po
Link to the Dataset(s)
PanNuke: https://jgamper.github.io/PanNukeDataset
CoNSeP: https://github.com/vqdang/hover_net
BibTex
@InProceedings{WuYon_Prompt_MICCAI2026,
author = { Wu, Yonghuang AND Liang, Zhenyang AND Zeng, Wenwen AND Xie, Xuan AND Yu, Jinhua},
title = { { Prompt Group-Aware Training for Robust Text-Guided Nuclei 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
The paper tackles the sensitivity of text-guided segmentation to prompt variations by viewing semantically similar prompts as a group and enforcing consistent predictions across them. Based on this idea, it introduces a simple group-aware training framework that combines quality-based weighting of prompts and a logit-level consistency constraint. The method does not modify the model architecture and can be easily applied to existing text-conditioned segmentation models. Experiments show improved robustness to prompt variations and better cross-dataset generalization.
- 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 addresses an important and practical problem in text-guided medical image segmentation, namely the sensitivity to prompt variations, which is particularly relevant in real clinical settings.
2.The key idea of modeling semantically equivalent prompts as groups and enforcing consistency is simple, intuitive, and well-motivated.
3.The proposed method is lightweight and easy to apply, as it does not require any architectural modification and only introduces additional training objectives.
4.The experimental setup is well-designed, including cross-dataset generalization and prompt quality variation, and the results consistently show improved robustness and reduced performance variance.
- 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.assuming semantic equivalence, which, while effective in controlled settings, may be less representative of real-world scenarios where prompt meanings are ambiguous or partially overlapping.
2.Prompt quality is defined based on segmentation loss using ground-truth masks, which is not available at inference time, raising questions about how prompt reliability is handled in practice.
3.Although the ablation study shows the benefit of the proposed design, it would be helpful to further disentangle how much of the gain comes from the group-aware objectives versus the use of multiple prompts during training.
4.The approach enforces output-level consistency while keeping the text encoder fixed, and therefore does not fundamentally improve the semantic understanding of prompts.
- Please rate the clarity and organization of this paper
Satisfactory
- Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.
The submission does not mention open access to source code or data, but provides a clear and detailed description of the algorithm to ensure 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 paper addresses an important and practical problem in text-guided medical segmentation, namely sensitivity to prompt variations. The proposed group-wise consistency formulation is simple and intuitive, and the method is easy to apply without architectural changes. The experiments are solid, showing consistent improvements across prompt quality levels and cross-dataset settings.
However, the methodological novelty is moderate, and the approach relies on predefined prompt grouping and ground-truth-based quality estimation, which may limit its applicability in more realistic scenarios. Overall, the paper is clearly presented and empirically well-supported, making it suitable for acceptance.
- Reviewer confidence
Very confident (4)
- [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.
N/A
- [Post rebuttal] Please justify your final decision from above.
N/A
Review #2
- Please describe the contribution of the paper
This paper reformulates prompt sensitivity in text-guided segmentation as a group-wise consistency problem and proposes a training-time framework to improve prompt robustness for nuclei 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.
1.The mapping from prompts to masks is a key clinical challenge in pathology, where different textual descriptions legitimately refer to the same target structure. The motivation of this work is therefore well-grounded.
2.The zero-shot generalization results are inspiring. The method achieves improvements on 5 out of 6 external datasets and reaches competitive performance with SAMPO using only text prompts.
- 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 optimization of L_group contradicts the stated motivation. The paper aims to make the model robust to low-quality prompts. However, L_group encourages the model to focus on high-quality prompts while relaxing on low-quality ones, effectively widening the gap. This directly conflicts with the goal of prompt invariance. The authors should explain how L_group and L_cons can work together when their optimization objectives are opposing.
2.A critical configuration is missing: removing only L_group while retaining L_cons. The current ablation cannot disentangle the independent contribution of L_group.
- The paper states that both q̃_i (Eq. 5) and the weights w_i (Eq. 6) are computed with stopped gradients. If both are detached from the computational graph, L_group contributes zero gradient to the model optimization. The authors should clarify precisely which quantities retain gradients and how L_group provides a training signal.
4.The paper layout needs refinement. Tables 3, 4, and 5 are placed after the Conclusion section, making them hard to follow when reading the main text.
- Please rate the clarity and organization of this paper
Satisfactory
- Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.
The submission does not provide sufficient information for reproducibility.
- Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?
N/A
- Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2026/en/REVIEWER-GUIDELINES.html
N/A
- Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.
(3) Weak Reject — marginally below the acceptance threshold, but would not mind if accepted, dependent on rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
See the above strengths and weaknesses.
- Reviewer confidence
Confident but not absolutely certain (3)
- [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
Review #3
- Please describe the contribution of the paper
This paper proposes a prompt group-aware training framework for text-guided nuclei segmentation. It models prompt variability by grouping semantically equivalent prompts and introduces quality-guided weighting and consistency regularization.
- 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 suggest that modeling variability as a population consistency problem is both simple and well-motivated.
- 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 relates to consistent learning and multi-view training. It focuses solely on prompt robustness and does not address the broader challenges in medical segmentation. It is currently unclear how cue groups are constructed in practice and whether the process is scalable. I’d like to know how sensitive this method is to misclassification. Can this method be extended to other modalities? What performance can this method achieve without a prompt? Moreover, some writing details to note: For example, it is not necessary to write the abstract in paragraphs.
- 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 mention open access to source code or data, but provides a clear and detailed description of the algorithm to ensure 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 score is based on a clear and well-motivated idea with consistent empirical gains. The method is simple and effective, improving robustness to prompt variations. However, the novelty is moderate and the scope is relatively limited. Overall, this leads to a weak accept recommendation.
- Reviewer confidence
Confident but not absolutely certain (3)
- [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 sincerely thank the reviewers and Area Chair for their careful evaluation of our manuscript and for the encouraging decision (Early Accepted). We are grateful for the recognition of the practical importance of prompt robustness in text-guided medical segmentation, and for the constructive suggestions that will help us further improve the final version. Below we summarize the main revisions we will make.
1.Scope and clinical relevance. Our work focuses on a controlled yet practically meaningful setting: multiple prompts that are intended to refer to the same segmentation target and therefore share the same supervision mask. We chose this setting to isolate a concrete reliability issue in text-guided segmentation, namely inconsistent predictions under alternative but clinically plausible phrasings. In the final version, we will state this scope more explicitly in the Introduction and Conclusion, and position broader prompt ambiguity as future work. 2.Training-only use of prompt quality. As noted by Reviewer 1, prompt quality is defined using segmentation loss with ground truth. We clarify that this signal is used only during training as a relative cue within each prompt group and introduces no extra requirement at inference time. Our inference procedure is unchanged: all methods are tested with a single prompt, without ensembling or online quality estimation. We will revise the method section to make this explicit. 3.Roles of
L_groupandL_cons. We agree that the relationship between these two losses needs clearer explanation. They are complementary rather than contradictory.L_groupprovides a relative optimization signal within a prompt group, allowing clearer prompts to serve as stronger anchors under the same supervision target, whileL_consdirectly encourages consistent predictions across prompts so that the model becomes less sensitive to wording variation. 4.Additional ablation and separation from multi-prompt training. We agree that it is important to distinguish the benefit of using multiple prompts from the benefit of the proposed group-aware objectives. The current results (table 2) already suggest that increasing prompt number (K) alone does not monotonically improve performance, and that grouped multi-prompt training without the proposed objectives is insufficient. Following Reviewer 2’s suggestion, we will add the missing ablation that removes onlyL_groupwhile retainingL_cons, so the individual contributions of the two losses can be isolated more cleanly. 5.Gradient-flow clarification. Reviewer 2 correctly identified an inconsistency in the manuscript description: if both the quality signal and weights were detached,L_groupwould yield no gradient. This is a writing error rather than an implementation issue. In the actual implementation, gradients are stopped only for the quality-derived ranking signal, while the weighting path remains differentiable. We will correct the equations and text so that the optimization description matches the implementation. 6.Prompt-group construction and extensibility. Prompt groups are constructed around the same image-target pair and shared mask, with prompts generated in a controlled manner by varying prompt type and prompt quality. This design does not require new mask annotations, only multiple semantically aligned descriptions for an existing target, which makes it practically scalable. We will make this pipeline clearer in the final version and release the code and prompt-construction details for reproducibility. We will also describe extension beyond pathology as a plausible future direction rather than a demonstrated result.We again sincerely thank the reviewers and Area Chair for their valuable feedback and for the positive decision. We will incorporate these suggestions carefully in the camera-ready version to further improve the clarity and quality of the paper.
Meta-Review
Meta-review #1
- Your recommendation
Provisional Accept
- 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.
I recommend provisional acceptance. The paper addresses a practical and timely problem in text-guided medical segmentation: robustness to prompt variation. The proposed group-aware training strategy is simple, lightweight, and easy to apply to existing models, and the experiments show consistent gains in prompt robustness and cross-dataset generalization. While the novelty is moderate and one reviewer raised important questions about the formulation of the group loss, gradient flow, and prompt-group construction, these issues appear clarifiable rather than fatal. Overall, the paper is well motivated and empirically convincing enough for acceptance, provided the authors address the methodological clarifications in the rebuttal or revision.