List of Papers Browse by Subject Areas Author List
Abstract
Volumetric Reasoning Segmentation (VRS) aims to segment a target region in a 3D medical scan from a free-form clinical query, where the referent is often implicit and requires both medical knowledge and volume-grounded reasoning. Existing methods typically rely on specialized segmentation tokens to connect language with mask decoding, but this coupling collapses the decision process into opaque latent representations, limiting interpretability and generalization to diverse narrative expressions. In this paper, we present MedVol-R1, a reinforcement learning-based framework for VRS that explicitly decouples evidence grounding from volumetric delineation: the LVLM grounds clinical reasoning to a verifiable 2D evidence anchor (key axial slice and 2D bounding boxes), which is then propagated into a coherent 3D mask by a frozen MedSAM2 module. We train MedVol-R1 with cold-start supervised fine-tuning followed by GRPO, guided by a multi-component reward that encourages informative evidence selection, accurate 2D spatial grounding, and cross-slice volumetric coherence, without requiring costly chain-of-thought annotations. Experiments on CT-ORG, AbdomenCT-1K, and KiTS23 from the M3D-Seg benchmark demonstrate that MedVol-R1 consistently outperforms strong baselines and achieves state-of-the-art performance, with reinforcement learning providing clear gains over pure supervised fine-tuning.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/2593_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to the Code Repository
https://github.com/AdvancingEther/MedVol-R1
Link to the Dataset(s)
M3D-Seg dataset: https://huggingface.co/datasets/GoodBaiBai88/M3D-Seg/tree/main
BibTex
@InProceedings{WanZic_MedVolR1_MICCAI2026,
author = { Wang, Zichun AND Shi, Hairong AND Wei, Bingzheng AND Xu, Yan AND Wang, Zihua},
title = { { MedVol-R1: Reward-Driven Evidence Grounding for Volumetric Reasoning 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 paper presents MedVol-R1, an innovative reinforcement learning-based framework for Volumetric Reasoning Segmentation. The core idea of explicitly decoupling clinical narrative reasoning from volumetric mask delineation is intuitive and effectively addresses the opaque latent representation issues of previous token-based approaches. Furthermore, the introduction of a multi-component reward system ensures genuine clinical understanding of the model.
- 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 framework explicitly decouples the language-driven reasoning process from the voxel-level 3D mask generation. Instead of relying on opaque, specialized segmentation tokens that collapse the decision process into an uninterpretable latent space, MedVol-R1 forces the LVLM to output a verifiable “2D evidence anchor”. A frozen MedSAM2 model then propagates this 2D anchor into a full 3D volumetric mask. This design solves the “black box” problem of existing LVLM segmentors. Furthermore, this is the first framework to apply GRPO to the task of VRS. It transitions the field away from pure SFT, which often leads to mere “surface-level pattern mimicry”. And the authors designed a highly tailored, multi-component reward system that evaluates the LVLM’s rollouts from four complementary perspectives.
- 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.Given the two-stage decoupled nature of the pipeline, the framework heavily relies on the LVLM to output a perfectly accurate 2D evidence anchor. If the LVLM reasons incorrectly, such as selecting the wrong key slice or inaccurate 2D boxes, can the frozen MedSAM2 identify, reject, or rectify this error? 2.The cross-slice consistency reward (Rc) evaluates the volumetric coherence strictly within a local axial neighborhood of ±Δ adjacent slices around the predicted key slice. The choice of this hyperparameter Δ is critical: if Δ is too small, it fails to enforce true global 3D consistency; if Δ is too large, MedSAM2’s slice-to-slice propagation might naturally degrade due to anatomical variations, unfairly penalizing the LVLM’s otherwise correct key-slice selection. The manuscript currently lacks an explanation or an ablation study on how Δ was chosen. 3.The experiments evaluate the framework on CT-ORG, AbdomenCT-1K, and KiTS23.While Figure 2 provides specific examples, the author should also clearly list the specific types of organs and lesions that the current method is capable of segmenting across these datasets. Furthermore, I am curious how the fixed anatomical labels were replaced with the “free-form linguistic descriptions” or implicit queries? 4.Figure 2 only provides qualitative comparisons between the GT, the M3D baseline, and MedVol-R1; the authors should include the visualization results of SAT and BiomedParseV2.
- 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?
My overall recommendation for this paper leans positive, acknowledging its significant conceptual novelty and strong empirical results, while recognizing several methodological and analytical gaps that should be addressed during the rebuttal phase.
- 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
novel 3d segemtnation framework with RL. predicting slide + 2d bounding box and then turning it into a 3d mask.
- 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.
auditable intermediate step. It aligns very well with the rationale on why they used RL. Ablation is very clean.
- 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.
They heavily rely on a single slice. And it is not very clear whether the performance gain came from their technical novelty or just giving MedSAM2 a better prompt.
- 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 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.
(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?
lack of reproduction
- 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 #3
- Please describe the contribution of the paper
This paper proposed the MedVol‑R1, a reinforcement learning-based framework for Volumetric Reasoning Segmentation (VRS) that explicitly decouples evidence grounding from volumetric delineation, addressing key limitations of existing methods that rely on opaque, specialised segmentation tokens. Unlike prior approaches that collapse the decision process into latent representations, MedVol‑R1 trains a Large Vision‑Language Model (LVLM) to first produce a verifiable 2D evidence anchor—a key axial slice and corresponding bounding boxes—which is then propagated into a coherent 3D mask by a frozen MedSAM2 model. This design enhances interpretability and generalisation, especially for free‑form, implicit clinical queries that require medical knowledge and spatial reasoning.
- 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) Apply GRPO (Group Relative Policy Optimization) to 3D medical volumetric reasoning segmentation, with a four-dimensional complementary reward (format, axial visibility, 2D IoU, cross-slice Dice), enabling structured reasoning without requiring CoT annotations. (2) Formalizes volumetric reasoning segmentation under implicit clinical queries as a distinct task, identifies the inherent limitations of existing “specialized segmentation token” paradigms in terms of interpretability and generalization, and proposes targeted improvements. The novel explicit decoupling of volumetric reasoning from geometric mask extraction effectively addresses the interpretability and generalization bottlenecks of current LVLM-based methods for implicit clinical queries. (3) Sufficiently comprehensive ablation studies, with quantitative validation of both the necessity of cold-start initialization and the contribution of each reward component. The results are self-consistent with the design motivations.
- 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) All experiments were conducted solely on three CT subsets of M3D-Seg. It is recommended to supplement with one or two MRI datasets or real-world clinical free-text queries that have not been paraphrased, in order to further validate cross-modal / real-world generalizability. (2) Source code should be open-sourced.
- 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?
No concern.
- 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) Abbreviation terms should be with the full name when it first exists in the paper (i.e., LMLM and GRPO in the abstract, SAM in the Introduction). (2) Why using Qwen3 in the experiments should be briefly introduced.
- 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 introduced a novel formulation by explicitly decoupling volumetric reasoning from geometric segmentation, replacing opaque specialised tokens with a verifiable two‑step pipeline (LVLM → 2D evidence anchor → frozen MedSAM2 → 3D mask). This is an application of Group Relative Policy Optimization (GRPO) to 3D medical volumetric reasoning segmentation, and the carefully designed four‑component reward function (format, axial visibility, 2D IoU, cross‑slice Dice) enables structured reasoning without costly chain‑of‑thought annotations. The evaluation is rigorous, comparing against two paradigms of baselines across three CT datasets. The method also demonstrates clinical feasibility on real CT volumes with a 4B Qwen model that runs on NVIDIA RTX A6000 GPUs. However, all experiments are limited to the M3D‑Seg benchmark; no validation on MRI, ultrasound, or other modalities is provided, leaving cross‑modal generalizability unproven. In addition, the source code is not open-sourced, hindering reproducibility and adoption by the research 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 sincerely thank all reviewers for their constructive comments and for recognizing the novelty of MedVol-R1.Q1(R1&R2): Is the gain from learned evidence grounding or merely from giving MedSAM2 a better prompt? What if the LVLM predicts a wrong anchor? A1: We thank the reviewers for this important question. MedSAM2 is essential but kept frozen, serving only as a low-level mask executor that delineates and propagates masks from a slice/box prompt. It does not understand the free-form VRS query, select the informative axial slice, or localize the query-relevant ROI; these decisions are learned by the LVLM through our evidence-grounding policy. Thus, the gain is not from improving MedSAM2 itself, but from learning to provide a query-conditioned, actionable 2D anchor.
We agree that a better prompt is crucial. However, the prompt is not manual or oracle-provided; it is predicted by the LVLM from the volume and query. Therefore, “giving MedSAM2 a better prompt” is precisely the effect of learned evidence grounding. Our ablations keep MedSAM2 fixed and vary the grounding strategy, isolating learned query-conditioned anchoring.
If the LVLM predicts a wrong anchor, MedSAM2 will usually propagate from the incorrect evidence, since it cannot semantically reject or correct a wrong prompt. To reduce such failures, we optimize axial visibility, 2D localization, and cross-slice consistency rewards. The explicit anchor also makes errors diagnosable as slice-selection, box-localization, or propagation failures.
We will clarify this limitation and discuss future agentic refinement, where the system inspects the initial 3D mask and refines it with MedSAM2’s positive/negative box or point prompts.
Q2(R1&R3): What targets are evaluated, and how are fixed labels converted into implicit queries? A2: For AbdomenCT-1K and CT-ORG, we follow M3D-Seg: targets come from original semantic labels, each mapped to multiple ChatGPT-generated descriptions in the M3D-Seg term dictionary. We randomly sample one as the implicit query, e.g., “liver” → “the organ that stores and regulates glycogen.” For KiTS23, we construct reasoning-oriented queries using mask-derived spatial/semantic relations, e.g., a cyst associated with the left kidney can be described as “an abnormal lesion on the left kidney.” We will clarify this protocol and show representative label–query pairs. Q3(R1): Choice of Δ in the cross-slice consistency reward. A3: We agree that Δ is a critical trade-off. We set Δ=5, forming an 11-slice neighborhood around the predicted key slice. Rc provides local rather than global volumetric supervision: too small Δ approaches 2D supervision, while too large Δ may introduce noisy rewards from anatomical variation and propagation errors. We will add the ablation study. Q4(R1): Qualitative comparison with SAT and BiomedParseV2.A4: Due to the page limit, Fig. 2 currently shows GT, M3D, and MedVol-R1.In the revision, we will include SAT and BiomedParseV2 on the same cases and annotate each with the label and sampled query. Q5(R2&R3): Reproducibility and code release. A5: We will release the training/inference code after the anonymity period. Q6(R3): Generalization beyond CT and M3D-Seg-style queries. A6: We agree that cross-modality validation and real clinical free-text queries are important. This work focuses on CT-based VRS because M3D-Seg subsets provide dense 3D annotations, public protocols, and controlled comparison with baselines, allowing us to isolate reward-driven evidence grounding. We will clarify that our claims are limited to CT-based VRS with M3D-Seg-style label-derived implicit queries; MRI/other modalities and uncurated clinical reports will be discussed as future work. Q7(R3): Minor revisions. A7: We will define all abbreviations at first mention, including LVLM, GRPO, and SAM, and clarify that Qwen3-VL-4B balances reasoning ability and computational feasibility for GRPO-based volumetric evidence grounding.
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 study explores a RL-based framework called MedVol-R1 for volumetric reasoning segmentation. It first trains a LVLM to produce a verifiable 2D evidence anchor, which is then propagated into 3D mask by MedSAM2.All reviewers agree that the framework is conceptually novel and the results are promising, and gives all positive rating. After reading the paper and reviews, AC is consistent with reviewers and decides to recommend an Early Accept. However, authors are encouraged to address the concerns raised by reviewers during the rebuttal.
