List of Papers Browse by Subject Areas Author List
Abstract
Accurate medical image segmentation requires both pixel-level precision and global anatomical consistency. However, the standard supervised paradigm, focused on pixel-wise distribution fitting, often traps models in sub-optimal local minima and fails to directly optimize non-differentiable structural metrics. In this paper, we propose Structural Group Relative Policy Optimization (S-GRPO), a novel paradigm that reformulates temporal sampling into structural parallel sampling. Departing from standard GRPO, S-GRPO leverages a multi-head architecture to generate a group of diverse segmentation candidates in a single forward pass, markedly reducing training latency. An adaptive Selector Network is designed to identify the optimal “expert” head and estimate relative advantages. Furthermore, we incorporate Advantage-Aware Distillation and a conditional Exponential Moving Average (EMA) update strategy to anchor the policy to a stable reference model, ensuring balanced exploration and exploitation. S-GRPO is designed as a plug-and-play optimization paradigm, compatible with diverse backbones including CNNs, Transformers, and State Space Models. Extensive experiments demonstrate that S-GRPO consistently improves performance across eight representative architectures, achieving superior results compared to conventional supervised baselines. The source code is available at: https://github.com/Duan-yiru/S-GRPO-main.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/1756_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to the Code Repository
https://github.com/Duan-yiru/S-GRPO-main.
Link to the Dataset(s)
N/A
BibTex
@InProceedings{DuaYir_SGRPO_MICCAI2026,
author = { Duan, Yiru AND Nie, Minxin AND Ma, Xiao AND Li, Xiaohui AND Zhang, Wenbo AND Chen, Qiang},
title = { { S-GRPO: Structural Group Relative Policy Optimization for Medical Image Segmentation } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 16883},
month = {September},
page = {pending}
}
Reviews
Review #1
- Please describe the contribution of the paper
The authors describe an approach for ‘augmenting’ standard medical image segmentation algorithms (the “backbones”) with a Reinforcement-Learning process, which allows the pipeline to improve performance over the selected baseline backbone. The rationale for this is that the RL component can better explore the state-space of possible ‘segmentors’, as well as introduce more learning components in the form of reward functions (e. g. to account for anatomical / boundary precision by also considering boundary metrics). Additionally they improve on state of the art by structuring their pipeline such that at each training iteration only one segmentor ‘head’ is updated rather than the entire collection, where the segmentor to be updated is chosen probabilistically on the basis of prior performance, and also defining a reference model as an exponentially weighted average over past policy model parameterisations, which is used to define the Advantage function and a loss function component, improving the overall stability of the RL process.
- 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 work represents a novel pipeline, building on state of the art and demonstrating improvements. The methodology is likely to lead to improvements in both accuracy and time-complexity over standard RL approaches (though obviously still likely more costly than the individual backbones).
- 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 is written in a somewhat dense and ‘messy’ manner; some explanations are not provided, while others appear far from the point of first use, etc. This makes reading a bit cumbersome, and having to read the paper in multiple passes in order to follow the proposed process properly.
-
Since one of the stated goals of the paper is improved time-complexity, and particularly with respect to GRPO, it would have been useful to compare performance (both in terms of accuracy as well as training time) against this as a benchrmark, and additionally, it would have been useful to compare training time against the backbone benchmark as well. However, training-time is not addressed anywhere in the paper’s results.
-
- 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
-
The paper lacks in explanations / exposition in many places. E. g. it would be useful to explain / clarify to the reader exactly the definition of the problem in standard RL terms, i. e. what is the nature of the states and actions on which the agent is acting; How is the Reference model initialised; What loss was used when the backbone was trained by itself for the purposes of comparison (and, e. g. , did this loss also take ASSD into account, and if not why is it a fair comparison?); Also some terms (e. g. ‘backbone’) are introduced without explanations (or explanations appear or are implied much later in the text) in contexts where such explanations would have been very useful.
-
Some notational terms are not properly defined, or there is abuse of notation making definitions unclear. E. g. what is C in \mathbb{R}^{C \times H \times W}? Channels? 3rd dimension? Count? What does \mathbb{D} = {(X,Y)} mean, did you mean something like { . . . }_{i=1}^N? The way this is written implies your dataset is a single image, which is confusing at first. What operation does Conv_k denote here (if it’s convolution, what is the kernel? if it’s the application of a CNN, what are the CNN’s important characteristics?).
-
Some areas could use more explicit insight in the text. For example:
1) why is GAP(F) useful, what does it actually do here beyond compressing the output to a single scalar? Why not choose an accuracy metric instead, for instance, given that the generated probabilities should reflect a measure of confidence?
2) Why combine Dice and Tanimoto (IoU) coefficients? Note that one can derive the Dice from the Tanimoto coefficient and vice versa (e. g. Dc=2Tc/(1+Tc) and equivalently Tc=Dc/(2-Dc). So Dc + Tc = (3Tc+Tc²)/(1+Tc) = (3DC-Dc²)(2-Dc). What do you gain from this? Is there some practical difference from combining these that can at least be shown empirically through ablation studies? Combining cross-entropy with Dice-loss further down is more reasonable, since the CE can mitigate the DL instability to some extent, but even this is done without explanation. It would be good to provide insight why you combine these, and if possible also reference relevant literature.
3) Regarding the distillation weight w_{kd}, the text states that when policy outperforms the reference, the weight (and thus KD-loss) decreases “to encourage exploration”, otherwise the reverse occurs “to enforce stability”. Can you explain this further? Shouldn’t it be the other way round? (i. e. shouldn’t unstable policies be restabilised, rather than further destabilised? and doesn’t reducing the loss stabilise rather than destabilise the model’s learning?).
4) JSD is an uncommon choice; why did you chose this particular divergence over other dissimilarity measures, e. g. something as simple as ||Mi-Mj||? (e. g. presumably one reason is because unlike other popular divergences the JSD is symmetric, so this would be a useful insight to add).
- Note that paper anonymity is broken somewhat in the provided code (see LICENSE and hardcoded paths in code). This looks accidental, but worth pointing out so that the authors are aware for future submissions.
-
- 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 is a novel contribution improving over state of the art, and would be of great interest to the MICCAI community.
- 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 proposes S-GRPO, a reinforcement-learning-inspired optimization framework for medical image segmentation. The key idea is to replace the temporal sampling in standard GRPO with structural parallel sampling, where multiple segmentation heads generate diverse hypotheses in a single forward pass, followed by selector-based routing, reward-guided advantage estimation, advantage-aware distillation, and conditional EMA stabilization. Experiments on ACDC and OCT across multiple backbone families show generally consistent improvements over the corresponding supervised baselines.
- 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.
Main Advantages of S-GRPO 1.This paper brings GRPO into dense pixel-level segmentation-a direction that, to our knowledge, has not been previously explored. By applying GRPO directly to pure segmentation models, S-GRPO stands as one of the earliest efforts in this space. 2.Instead of relying on GRPO’s expensive multi-round iterative sampling, this work introduces a multi-head architecture that achieves the same goal in a single forward pass. The design is intuitive, computationally efficient, and naturally fits the high-resolution requirements of medical imaging. 3.On the experimental side, S-GRPO proves to be a general-purpose optimization booster that works across different backbones, rather than a trick tailored to any specific architecture.
- 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.
Major Concerns: 1.The selector operates on GAP-compressed shared features rather than directly observing the candidate segmentation hypotheses. Since global average pooling removes substantial spatial and boundary information, it is unclear whether the selector can truly identify the structurally best head for a given sample, or whether it is merely learning a coarse sample-level prior. This is particularly important because the quality differences among candidate heads are presumably determined by fine-grained structural properties, which may not be well preserved after GAP. 2.The paper does not provide sufficient evidence that the multiple heads indeed learn meaningfully distinct structural hypotheses. There is no ablation on the number of heads, so it is unclear how sensitive the method is to this design choice or whether the selected value is empirically justified. In addition, the paper does not visualize or quantify whether different heads focus on different structures or failure modes, rather than behaving as weakly perturbed replicas of one another. 3.The experimental validation is somewhat limited. Since the paper positions S-GRPO as an RL-based optimization paradigm, comparing only each backbone with and without S-GRPO is not fully sufficient. It would be more convincing to include representative alternative optimization baselines, such as reward-driven or other RL-inspired segmentation methods, to better justify the necessity of the proposed formulation.
Minor Concerns: 4.The result analysis is not yet sufficient. In Table 1, the gains on ACDC for CNN backbones are relatively small (e.g., U-Net and U-KAN), compared with the larger improvements on several Transformer/Hybrid/Mamba backbones. The paper should provide a clearer analysis of this backbone-dependent behavior. 5.The paper should provide a clearer analysis of computational cost. Since S-GRPO introduces multiple parallel heads, a selector, a reference model, and extra optimization terms, reporting only the performance gains is not fully sufficient. Quantitative comparisons of training time, memory usage, and parameter/FLOPs overhead would help justify the claimed efficiency and plug-and-play practicality.
- 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 proposed S-GRPO addresses a timely and underexplored problem by migrating GRPO from LLMs to dense segmentation, and the structural parallel sampling method consistently improves eight diverse architectures. Its value lies more in opening a new paradigm than in technical depth. However, a critical ablation experiment is missing to disentangle the contribution of reinforcement learning from that of the ensemble strategy itself, and no statistical significance testing is provided.
- 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
The main contributions of the proposed work include (i) Novel framework based on structural parallel sampling for dense prediction (ii) Coupling of expert head mechanism with dynamic selection and (iii) S-GRPO as an aid for effective segmentation in multiple architectures
- 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 manuscript’s strength lies at the multiple checkpoint strategies at each stage. At sampling stage, k parallel segmentation heads were used with back propagation as selection strategy. Reward modeling is designated to be hybrid to ensure accuracy at each pixel and boundary. Multi-objective optimization is proposed to ensure beneficial exploration.
- 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.
Given that GRPO is a widely adopted for similar research problems, providing a comparative analysis is crucial. To better highlight the impact of the research, I suggest adding a comparison between S-GRPO and the generic GRPO. Evaluating both across the various architectures mentioned would provide a clearer picture of how much efficiency is gained through structural parallel sampling.
- 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
The authors are suggested to cite appropriate references for methods like Advantage-Aware Distillation that are already available in the existing 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.
(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?
The manuscript overall is well written and it has technical merit to be accepted. The result section could be made impactful by including performance benchmarking with existing GRPO.
- 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
Author Feedback
N/A
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.
This work has received overwhelmingly positive reviews from all assigned reviewers. Those reviews consistently recognized its novel pipeline of medical image segmentation with reinforcement learning process, experimental improvements, and well-written. While the work is of good quality, the reviewers have raised minor concerns that authors are advised to address. Based on the good quality and novelty, this work is recommended for early acceptance.
