List of Papers Browse by Subject Areas Author List
Abstract
Counterfactual medical image generation enables what-if'' clinical exploration through instruction-conditioned editing. However, existing methods often lack the multimodal reasoning required to align abstract directives with precise visual modifications, failing to balance instruction compliance with anatomical preservation. Reinforcement learning offers a promising solution, but it remains largely unexplored in this field. In this paper, we propose MedXEdit-GRPO, a reinforcement learning framework for counterfactual medical image editing. To provide stable and comprehensive feedback, we curate MedXEdit-Reward, a large-scale dataset of 30,000 samples with rationales, and develop MedXEdit-Score, a reward model enforcing areasoning-before-scoring’’ paradigm for semantically grounded optimization. We optimize our generative policy via Group Relative Policy Optimization (GRPO), which effectively harmonizes diverse clinical objectives into a unified, stable training process. Quantitative and qualitative evaluations demonstrate that MedXEdit-GRPO significantly outperforms state-of-the-art methods in both instruction compliance and anatomical consistency, highlighting its potential for robust and interpretable medical analysis.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/0241_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to the Code Repository
https://github.com/MedXEdit-GRPO/MedXEdit-GRPO
Link to the Dataset(s)
https://github.com/MedXEdit-GRPO/MedXEdit-GRPO
BibTex
@InProceedings{PanYan_MedXEditGRPO_MICCAI2026,
author = { Pan, Yaning AND Xiao, Renwei AND Li, Qingqiu AND Li, Xiaoyi AND Li, Xiaoling AND Feng, Rui AND Zhang, Xiaobo},
title = { { MedXEdit-GRPO: Reasoning-Aware Online RL for Counterfactual Medical Image Generation } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 16878},
month = {September},
page = {pending}
}
Reviews
Review #1
- Please describe the contribution of the paper
They create a dataset for counterfactual medical image generation and use GRPO to train realistic counterfactual image generation.
- 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 uses GRPO to train a model that performs counterfactual edits on chest X-ray images. Additionally, the authors claim to release the dataset, code, and model weights.
- 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 tuple (P, I_in, I_out) or (P, I_in, I_target) is mentioned in Section 3.1 without a description of what the variables mean. There is only a partial explanation of the variables in Section 3.2. You say you train on 24 GPUs, please also provide the amount of VRAM per GPU to give a better impression of the needed resources. No statistical evaluation is performed. Confidence intervals, e. g. based on bootstrapping, would be helpful to better assess the differences in the results. The word “significantly” is used without showing statistical significance. Perhaps “substantially” or similar would be better, so that the reader does not get a wrong impression (or alternatively, the authors could add a test for statistical significance). In the first sentence of Section 3.1, it says “Inspired by [28,16]”, however these two papers are not mentioned in the Related Work section. It is not described in much detail how the images for the MedXEdit-Score are generated.
- 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 authors claimed to release the source code and/or dataset upon acceptance of the submission.
- 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 idea of the paper is interesting and the associated dataset could be valuable for other studies. However, a statistical evaluation / assessment of the results is missing.
- 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 #2
- Please describe the contribution of the paper
The paper introduces a complete Reinforcement Learning (RL) framework for counterfactual medical image generation (specifically Chest X-rays). The main contributions are: 1.MedXEdit-Reward: A dataset of 30,000 cases containing source/edited images, instructions, rationales, and scores, distilled using Gemini-3-Flash. 2.MedXEdit-Score: A specialized reward model fine-tuned on the Qwen3-VL-4B backbone that employs a “reasoning-before-scoring” paradigm to provide semantically grounded feedback. 3.MedXEdit-GRPO: An online RL optimization pipeline that utilizes Group Relative Policy Optimization (GRPO) to fine-tune a medical generative model (initialized from UniMedVL) to improve instruction compliance and anatomical consistency.
- 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.Transitioning to an online RL framework (GRPO) for medical image editing is original. 2.By forcing the reward model to generate a rationale before a scalar score, the authors leverage the VLM’s internal logic, which significantly improves the reliability and interpretability of the reward signal. 3.The authors tackle the entire RLHF pipeline, from the curation of a novel preference dataset to the training of the reward model, and finally the policy optimization.
- 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 entire reward model relies on the distillation of rationales and scores from Gemini-3-Flash. While Gemini is a powerful general-purpose VLM, medical image interpretation (especially subtle counterfactual changes in CXRs) requires deep domain expertise. Although the authors mention a 200-pair verification by radiologists, 200 is a relatively small sample size compared to the 30,000 training samples. The paper lacks a deeper analysis of potential hallucinations, biases, or systematic errors introduced by the Gemini teacher model. 2.Since the generative policy is initialized with UniMedVL, UniMedVL serves as the most critical baseline. The authors should clearly state the improvement compared to this baseline. 3.Table 3 compares MedXEdit-Score against a “Hybrid Reward”. However, a more insightful ablation would isolate the reasoning component during the RL phase. For instance, how does the RL policy perform if trained with a reward model that only outputs scores without the reasoning step?
- 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 authors claimed to release the source code and/or dataset upon acceptance of the submission.
- 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
1.I find the choice of GRPO reasonable, but there is no comparison against other standard alignment algorithms used in image generation (e.g., DPO, PPO). 2.In Section 1, “MedEdit-Reward” and “MedEdit-Score” are used in the contribution list, but elsewhere they are referred to as “MedXEdit-Reward” and “MedXEdit-Score”. Please ensure consistent naming.
- 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 presents a novel and technically sound application of RL (GRPO with a reasoning-based reward model) to counterfactual medical image editing. However, concerns remain regarding the reliance on a general-purpose VLM (Gemini-3-Flash) for curating the 30k medical dataset, the inadequate discussion of the UniMedVL baseline, and the limited ablation studies.
- 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 authors present MedXEdit-GRPO, a reinforcement learning-based framework designed for instruction-conditioned counterfactual editing of chest X-rays. The paper outlines three primary contributions: (1) MedXEdit-Reward, a decent-sized reward dataset; (2) MedXEdit-Score, a “reasoning-before-scoring” reward model; and (3) the application of GRPO-based policy optimization to an image editor. Empirically, the proposed method demonstrates strong overall performance against established baselines across several metrics, including AUC, CLIP-I, FID, etc. with a custom holistic score.
While the paper claims significant conceptual novelty, the broader literature already features robust works in medical counterfactual image editing (e.g., ProgEmu and InstructX2X at MICCAI 2025, RadEdit at ECCV 2024) and RL-based visual editing (e.g., HIVE at CVPR 2024, CXRL at MICCAI 2024). Consequently, the work’s primary contribution lies in the specific, effective integration of a reward-model-based online RL pipeline into counterfactual chest X-ray editing, rather than introducing a fundamentally new algorithmic paradigm for image generation.
- 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.
End-to-end design: authors construct a complete, cohesive pipeline encompassing reward data generation, reward modeling, and RL fine-tuning. This full-stack approach makes the system practically applicable and thoroughly realized.
Reward modeling: The “reasoning-before-scoring” framework is well-considered. The proposed reward model outperforms generic medical baselines as well as a non-reasoning variant, validating that the evaluator is executing a nontrivial and beneficial reasoning step.
Empirical performance: The proposed method achieves state-of-the-art results across multiple key metrics when compared directly to a strong set of established baselines.
Clinical relevance: Counterfactual exploration and robustness analysis are meaningful targets in medical imaging. Qualitative results show that the model’s edits are appropriately localized. Furthermore, shifting the field from SFT-oriented approaches toward reward-guided policy alignment represents a valuable progression for the domain.
- 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.
Overstated novelty: The manuscript frames reinforcement learning as a largely unexplored avenue in this domain, which overlooks adjacent precedents. For example, CXRL utilized policy-gradient RL for chest X-ray generation, and HIVE employed human-feedback reward modeling for general image editing. Given that existing works already map much of the task landscape, the contribution should be framed as a strong domain-specific integration rather than a foundational breakthrough.
Reward model bias: Distilling reward data primarily from Gemini-3-Flash and relying on only three source generators introduces a substantial risk. The learned reward model may easily inherit the specific biases, preferences, and blind spots of this narrow set of teachers.
Limited dataset contribution: The dataset curation relies heavily on the automated extension of open datasets using AI tools. With only 200 radiologist-verified preference pairs, the dataset lacks the extensive expert curation required to be considered a major standalone contribution, as the current automated methodology is relatively trivial to replicate.
Evaluation circularity: Employing MedXEdit-Score as both the reward function for RL and the final “Holistic Score” metric creates a circular evaluation loop. This diminishes the persuasiveness of the reported holistic gains, making the external, independent metrics far more reliable indicators of success.
Narrow validation scope: The experiments are exclusively conducted on chest X-rays derived from ICG-CXR. There is no external dataset validation, nor is there evidence demonstrating generalization to other modalities or distinct CXR distributions.
Insufficient expert evaluation: While radiologists verified a small subset of preference pairs, the paper lacks a comprehensive expert study evaluating the final edited images. Consequently, claims regarding clinical plausibility and anatomical preservation rely almost entirely on proxy metrics rather than direct clinical judgment.
Incremental performance gains: Although the method outperforms strong baselines, the improvements are modest, and the framework does not achieve universal superiority across all evaluated metrics.
- 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 authors claimed to release the source code and/or dataset upon acceptance of the submission.
- 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?
This paper tackles an important and timely problem in medical imaging with a well-engineered, technically cohesive framework. The proposed reward model is a conceptually sound component, and the empirical results consistently demonstrate favorable performance against strong baselines. These factors drive my recommendation for acceptance.
However, the conceptual novelty is moderate when viewed alongside recent advancements in venues like MICCAI and CVPR. The submission is primarily constrained by its limited human expert validation, potential circularity in its evaluation metrics, and a narrow experimental scope confined to a single modality.
Ultimately, this is a practically useful paper backed by strong execution. Despite the incremental novelty and validation limitations, the overall quality, relevance, and cohesive system design meet the bar 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
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.
The paper presents a novel and technically sound application of RL on medical image generation, and all reviewers aggree to accept.