List of Papers Browse by Subject Areas Author List
Abstract
While Vision-Language Models (VLMs) show great promise in volumetric medical report generation, they frequently suffer from visual hallucinations and a lack of grounding in 3D CT data. Current Supervised Fine-Tuning (SFT) and Reinforcement Learning (RL) strategies typically optimize text fidelity alone, essentially rewarding correct diagnoses derived from language priors rather than genuine visual perception. To address this, we propose cross-view aligned Evidence-driven Multimodal Reinforcement Learning (Evidence-MRL, noted as E-MRL), a reliable RL reasoning framework that formulates the generation process as a Markov Decision Process of “diagnosis-localization-verification”. Unlike standard approaches, our model is explicitly trained to identify a “key evidence slice” alongside the global diagnostic report, grounding its findings in verifiable visual evidence. Crucially, we introduce a novel cross-view consistency reward, which validates the semantic alignment between the golden-standard report and a local visual re-query of the selected key slice, providing additional rewards for correctly-localized reasoning. Experiments on large-scale 3D CT tumor datasets demonstrate that E-MRL significantly reduces hallucinations and improves diagnostic accuracy compared to SFT and RL baselines, offering a clinically interpretable solution for visually-grounded and tumor analysis.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/0536_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to the Code Repository
N/A
Link to the Dataset(s)
N/A
BibTex
@InProceedings{LiSij_EMRL_MICCAI2026,
author = { Li, Sijing AND Qiu, Zhongwei AND Wang, Zhuoya AND Yun, Boxiang AND Yi, Zhenyu AND Xu, Jianwei AND Zhang, Wenqiao AND Xia, Yingda AND Zhang, Ling},
title = { { E-MRL: Cross-View Aligned Evidence-Driven Multimodal Reinforcement Learning for Reliable 3D Tumor Analysis } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 16882},
month = {September},
page = {pending}
}
Reviews
Review #1
- Please describe the contribution of the paper
1.This paper addresses the critical issue of visual hallucination and lack of grounding in 3D medical vision-language models, particularly for tumor analysis in volumetric CT data. 2.The authors propose an evidence-driven multimodal reinforcement learning framework (E-MRL), which decomposes the task into a diagnosis–localization–verification process to improve interpretability and reliability. 3.A multi-granularity reward design is introduced, combining tumor existence, fine-grained attribute prediction (e. g. , size, location, enhancement), and cross-view consistency between global reports and local slice-level evidence. 4.The method incorporates an explicit visual grounding mechanism by predicting a key evidence slice and verifying it through a local re-query process, encouraging alignment between generated text and visual evidence. 5.Extensive experiments on 3D tumor datasets demonstrate improved diagnostic accuracy and better visual grounding compared to SFT and RL-based 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.
1.This paper investigates the critical issues of visual hallucination and lack of visual grounding in medical vision-language models for 3D tumor analysis and clinical applications. 2.The framework design is clear and intuitive. The proposed diagnosis–localization–verification pipeline is easy to understand and aligns well with the workflow of radiologists, improving interpretability compared to purely end-to-end text generation models. 3.The method strengthens visual grounding by requiring the model to predict a key evidence slice and verify it through a local re-query process, thereby linking textual predictions to visual evidence. 4.The combination of tumor existence, fine-grained attribute supervision, and cross-view consistency provides richer training signals than standard text-based objectives. 5.The consistency between the global report and local slice-level evidence offers a reasonable way to reduce hallucination and improve reliability. 6.The method is compared with multiple strong baselines and demonstrates consistent improvements in both diagnostic accuracy and attribute-level reliability.
- 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.Why is reinforcement learning used? What is the necessity of applying RL in this task? 2.What are the main contributions of the proposed method for this task? 3.Are there no prior works addressing this task? 4.What is the difference between using reinforcement learning and supervised fine-tuning (SFT) for this task? 5.What are the GPU requirements and training time for this method? 6.Does the method overly rely on LLM-based evaluation for reward design? 7.How does reinforcement learning mitigate hallucinations?
- 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?
I believe the main reasons are that the experimental task is relatively complete, the paper compares a wide range of methods, and the reward design shows a certain level of novelty. In addition, the following aspects also support my rating:
1.The proposed framework is well-structured, and the diagnose–localize–verify pipeline is intuitive and aligns with real clinical workflows. 2.The method emphasizes grounding by linking textual predictions with visual evidence, which improves interpretability. 3.The use of multiple supervision signals (e.g., tumor existence, fine-grained attributes, and cross-view consistency) provides richer reward signals. 4.The evaluation is relatively comprehensive, including both diagnostic accuracy and attribute-level reliability, making the results more convincing. 5.The method demonstrates consistent improvements over strong baselines, including both academic and commercial models.
- 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
Review #2
- Please describe the contribution of the paper
1.The paper proposed a novel “Diagnosis-Localization-Verification” RL framework for fine-grained evidence-driven reasoning.
2.The paper designed a comprehensive reward mechanism that integrates textual diagnostic correctness with visual evidence grounding.
3.The paper developed a global-to-local alignment mechanism that synchronizes global diagnostic findings with local slice level evidence.
- 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 aims to address an important failure mode of medical VLMs: generating correct-sounding reports based on language priors rather than actual visual evidence. The analogy to radiologist workflow is reasonable. The cross-view consistency alignment is also intuitive and reasonable.
- 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 SFT process lacks sufficient explanation. As shown in Tables 1 and 2, the most substantial performance gains appear to stem from the SFT stage. However, neither the training corpus nor its construction process is clearly described in the paper. A more detailed explanation of these components is necessary to evaluate the validity and reproducibility of the proposed approach.
2.The use of the “3D” concept in this paper warrants further clarification. Although the model takes CT scans as input by treating them as video sequences, this type of input arguably constitutes a 2D+1D representation, as the information is derived solely from the axial view. The authors should clarify whether this truly qualifies as 3D processing and address this distinction more explicitly.
3.Several existing works have similarly employed lesion localization methods to enhance MLLMs’ lesion-grounded reasoning and diagnostic capabilities [1][2]. The paper would benefit from a more thorough discussion of how the proposed method differs from and compares to these CT-grounded reasoning models, in order to better contextualize the contributions and provide a more comprehensive analysis.
[1] Wu, Linshan, Jiaxin Zhuang, and Hao Chen. “Glance and Focus Reinforcement for Pan-cancer Screening.” arXiv preprint arXiv:2601.19103 (2026).
[2] Li, Yifan, et al. “MedReason-R1: Learning to Reason for CT Diagnosis with Reinforcement Learning and Local Zoom.” arXiv preprint arXiv:2510.19626 (2025).
- 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
From Fig.2, it seems that the bounding box is missing from the generated report. Further explanations for this missing part would be nice for readers to understand the generated report, since the bounding box output is a key element in the framework.
- 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 research problem and motivation are both reasonable and valuable. However, the paper would benefit from a more thorough discussion and comparison between the proposed method and the most closely related lines of work. Additionally, the experiments regarding the SFT stage, which appears to account for the largest performance gain in the framework, require further clarification.
- Reviewer confidence
Confident but not absolutely certain (3)
- [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.
I thank the reviewer for their detailed response. The clarifications provided regarding the SFT procedure and implementation details have largely addressed my concerns. For the final version, I encourage the authors to incorporate these implementation details more explicitly into the manuscript, strengthen the comparison with prior methods, and revise the figures.
Review #3
- Please describe the contribution of the paper
The paper proposes E-MRL, a multimodal RL framework for 3D CT tumor report generation that encourages visual grounding by requiring the VLM to output both a diagnostic report and a key evidence slice. The selected slice is re-examined by a verifier model, and the resulting local description is used to construct a cross-view consistency reward together with tumor existence and fine-grained lesion attribute rewards. The main contribution is therefore an application-specific reward design that links global report generation with local slice-level visual evidence, aiming to reduce hallucinations and improve interpretability in 3D tumor analysis.
- 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 tackles an important and clinically relevant problem: visually grounded 3D CT tumor report generation. A key strength is the evidence-driven design, where the model is required to generate both a global report and a key evidence slice, followed by local slice re-query for verification. This provides a more interpretable and visually grounded alternative to standard text-only report optimization. The multi-granularity reward, combining tumor existence, lesion attributes, and cross-view/local evidence consistency, is well aligned with clinically meaningful reporting requirements. The evaluation is also relatively comprehensive, covering commercial VLMs, 2D and 3D medical VLMs, RadGPT, SFT baselines, reward ablations, and both diagnostic and visual-grounding metrics.
- 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.Limited algorithmic novelty. The MDP framing and GRPO-style process optimization are not fundamentally new; autoregressive VLM generation can already be cast as an MDP, and prior work has explored GRPO/process supervision and medical VLM RL. The more specific contribution is an application-driven combination of key-slice localization, local re-query, and reward design, rather than a new RL framework. The paper itself cites related medical VLM/RL or grounding works such as MedVLM-R1, Med3D-R1, VALOR, and MedMO.
2.Evaluation has potential circularity and judge-dependence. Fine-grained rewards and metrics rely on Qwen-235B judging size, location, and enhancement. In addition, the authors state that Table 2 uses RadGPT-generated fine-grained reports as references, which weakens the clinical validity of the attribute and grounding evaluation.
3.The RL gain over SFT is moderate, and baseline fairness is unclear. The largest improvement comes from SFT: average B-Acc increases from 53.55% to 83.57%, while E-MRL further improves it to 87.93%. Thus, the specific contribution of the proposed RL component is smaller than the overall headline gain. It is also unclear whether commercial and 2D VLM baselines receive comparable 3D visual inputs.
4.Reproducibility and presentation issues. Important details are missing, including GRPO group size, update steps, KL/clipping hyperparameters, judge prompts, key-slice parsing failures, and 3D volume preprocessing. The writing and formatting also need polishing, with grammar issues, inconsistent quotation marks/fonts, and typos.
- 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.
(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?
I recommend weak reject. The paper addresses an important MICCAI relevant problem: grounded 3D CT tumor report generation. The key slice localization and local re query idea is clinically motivated. However, the algorithmic novelty is limited. The main contribution is an application specific combination of MDP/GRPO, key slice localization, and multimodal rewards rather than a fundamentally new RL framework. I also have concerns about the evaluation. Fine grained rewards and metrics rely on Qwen 235B judging, and Table 2 uses RadGPT generated reports as references, raising judge dependence and circularity concerns. In addition, the gain from E MRL over SFT is moderate compared with the large improvement from SFT itself, and several reproducibility details are missing. The writing and formatting also need substantial polishing. These issues led me to a weak reject recommendation.
- Reviewer confidence
Confident but not absolutely certain (3)
- [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 has adequately addressed my concerns. The introduction of a verification stage is novel and useful for the task.
Review #4
- Please describe the contribution of the paper
This paper proposes E-MRL, a multimodal reinforcement learning framework for reliable 3D CT tumor analysis using vision-language models. The method reformulates report generation as a “diagnosis–localization–verification” process, where the model not only generates a global diagnostic report but also selects a key evidence slice for verification. A multi-granularity reward is introduced, combining tumor existence, fine-grained attribute correctness, and a novel cross-view consistency reward that aligns global report content with local visual evidence. Experiments on the AbdomenAtlas3.0 dataset demonstrate improved diagnostic accuracy, attribute grounding, and reduced hallucination compared to SFT and RL 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.
1) The paper addresses a critical issue—lack of visual grounding and hallucination in medical report generation—which is highly relevant and important for clinical deployment.
2) The “diagnosis–localization–verification” paradigm is intuitive and well aligned with radiologist workflows.
3) The idea of: first predicting a key slice then re-querying it and enforcing consistency with global report is conceptually strong and directly targets hallucination.
- 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) Fine-grained attribute rewards and evaluation rely on Qwen-235B, which may introduce bias and uncertainty in scoring reliability.
2) The “ground truth” key slice is derived from segmentation masks (max tumor area), which may not fully reflect clinically meaningful evidence selection.
3) Limited analysis of failure cases and 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
Consider citing several new works and use RFT + SFT at the same time in Report Generation, i.e. not only in 3D CTs.
- 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 paper tackles an important and timely problem in medical vision-language modeling by addressing hallucination and lack of visual grounding. The proposed diagnosis–localization–verification framework, together with the cross-view consistency reward, introduces a meaningful and well-motivated paradigm for evidence-driven reasoning. The method is technically sound, clearly presented, and supported by strong empirical improvements over competitive baselines. While some aspects such as reliance on LLM-based evaluation and computational cost could be further analyzed, the overall contribution is significant and relevant to the MICCAI community.
- 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 thank the reviewers for the constructive feedback. We respectly respond as follows. @R1-Q2, R4-Q1 Main contribution. We clarify that our key innovation is not a general policy optimization algorithm, but a diagnosis-localization-verification reasoning framework for CT tumor report generation. The model first performs a global diagnosis on the full CT and predicts a key slice for local reasoning. And we introduce a novel cross-view consistency reward to encourage reliable evidence localization. Accordingly, the main contributions are: (1) extend one-shot report generation to two-stage reasoning. (2) enforce global-local consistency via key-slice re-query. (3) shift the optimization target from text to evidence-grounded correctness, mitigating hallucination.
@R2-Q1, R4-Q3, R1-Q1457 The distinct roles and implementation details of SFT and RL. Our results show that general-purpose models(e.g. Qwen, commercial models), perform poorly on tumor tasks, making SFT necessary for domain-specific initialization and explaining the gain from 53.55 to 83.57.However, SFT mainly fits the text distribution and lacks interpretability and reliability. Our evidence-based RL improves lesion-attribute correctness and evidence-grounding, which are more clinically relevant to radiologists. On attribute VQA(Tab 2), our E-MRL improves over SFT from 57.49 to 67.53(+17.5%). In implementation, we built ~9.2K single-organ SFT samples from AbdomenAtlas3.0 with mask-derived key-slice supervision. For RL, we select about 2K hard cases with incorrect attribute predictions for targeted optimization. Training on 16 H20 GPUs takes about 12h for SFT and 14h for RL.
@R1-Q3, R2-Q3, R4-Q1 Related work and our novelty. Thank you for suggestions and we will add the missing citations. The key difference of our work is that we use VLM’s intermediate reasoning to localize the key slice, rather than relying on preprocessed suspicious regions(e.g. segmentation masks). Moreover, our rewards supervise the outcome, reasoning process, and verification simultaneously. Most prior methods either lack explicit grounding(e.g. MedVLM-R1, Med3D-R1, VALOR), or rely on preprocessed suspicious regions as auxiliary input(e.g. MedReason-R1). GF-Screen optimizes pan-cancer screening, which differs from our task.
@R1-Q6, R4-Q2, R5-Q1 Evaluation protocol, Qwen-235B judge, and RadGPT references. There is no evaluation circularity and the GT annotations are clinically rigorous. Qwen-235B is used only as a tumor-attribute parser(e.g., size, location, enhancement), not a direct evaluator. In Tab 1, sen., spec., and B-Acc are computed from parsed tumor existence. In Tab 2, the same parser is applied to both model outputs and GT reports to extract attributes for fair comparison. In addition, a radiologist on our team manually reviewed 10% of test set and confirmed 100% agreement with the parsed attributes. Regarding R4’s concern about RadGPT reference, the original RadGPT paper(ICCV2025) states that the GT, i.e., “RadGPT-generated report”, was revised and confirmed by 12 board-certified radiologists, ensuring high clinical quality.
@R2-Q2, R4-Q4 3D terminology, reproducibility details, and presentation. Our task is defined on 3D volumetric CT, while the input is formatted as a video sequence to match backbone. We will clarify this wording to avoid confusion. We will also add more implementation details. Concretely, our RL uses GRPO for 5 epochs with AdamW(lr 5e-6), gradient accumulation 8, group size 8, and KL coefficient β = 0.01.Full preprocessing and model configurations will be released in code repository.
@R5-Q2Q3 Failure mode and key slice selection. We analyzed the failure cases and found that common failure cases include very small tumors(<1 cm), multiple co-existing tumors, and abnormal enhancement under poor CT quality. We also acknowledge possible annotation variance in mask-based key-slice supervision and will add this discussion in the final version.
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 paper recieved mixed ratings. The authors should rebut on methodology contribution and clarification on the implementation details from R2 and R4.
- 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 authors significantly clarify the implementation details and the novelty of their work. I vote for accept considering its methodology novelty and the general ratings. VLMs for 3D tumor data would likely to raise discussion at MICCAI 2026.
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 authors have addressed the reviewers’ concerns well and backed their responses with evidence. While some limitations remain, these can be handled through clarifications, along with suggested corrections, in the final manuscript.