List of Papers Browse by Subject Areas Author List
Abstract
Developing 3D vision-language models (VLMs) with robust clinical reasoning remains challenging due to the high-dimensional complexity of volumetric imaging and biases in clinical narratives. We propose Med3D-R1, a unified framework that enhances diagnostic accuracy and interpretability for 3D CT-based abnormality diagnosis. First, we identify a structural narrative bias in standardized radiology reports and introduce an Abnormality Re-Weighting (ARW) strategy to counteract positional priors, together with a Residual Alignment Mechanism (RAM) that stabilizes cross-modal mapping between volumetric features and textual embeddings. Furthermore, we implement a reasoning-aware reinforcement learning stage with a consistency reward that encourages intermediate reasoning traces to align with report-supported evidence under report-based supervision. Evaluated on CT-RATE and the external RAD-ChestCT Multiple-choice Visual Question Answer benchmark, Med3D-R1 achieves state-of-the-art performance and improves diagnostic consistency and reasoning transparency in 3D medical VLMs. Code is available at: https://github.com/laihaoran/Med3D-R1.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/0550_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to the Code Repository
https://github.com/laihaoran/Med3D-R1
Link to the Dataset(s)
N/A
BibTex
@InProceedings{LaiHao_Med3DR1_MICCAI2026,
author = { Lai, Haoran AND Jiang, Zihang AND Zhang, Kun AND Yao, Qingsong AND Wang, Rongsheng AND He, Zhiyang AND Tao, Xiaodong AND Wei, Wei AND Zhou, Shaohua Kevin},
title = { { Med3D-R1: Mitigating Narrative Bias and Enhancing Reasoning Consistency in 3D Medical Vision-Language Models } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 16879},
month = {September},
page = {pending}
}
Reviews
Review #1
- Please describe the contribution of the paper
This paper addresses three critical challenges in 3D medical Vision-Language Models (VLMs) for CT diagnosis, including unstable alignment of high-dimensional 3D features with sparse text, structural narrative bias in radiology reports leading to “prediction inertia” and reasoning hallucinations. The authors propose Med3D-R1, a framework with three key contributions: 1) Residual Alignment Mechanism (RAM) is introduced to stabilize 3D-to-text projection via residual interpolation toward a fixed semantic anchor. 2) Abnormality Re-Weighting (ARW) identifies and mitigates positional bias where normal findings precede abnormalities in structured reports, improving sensitivity to pathological findings through token-level loss re-weighting. 3) Reasoning-Aware Reinforcement Learning introduces a consistency reward that aligns the model’s intermediate reasoning
with evidence from reference reports, thereby reducing clinically implausible hallucinations. Experiments on CT-RATE and RAD-ChestCT demonstrate state-of-the-art performance and improved reasoning transparency compared to existing 2D and 3D VLMs.
- 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 is the first to explicitly identify and quantify the structural narrative bias in standardized radiology reports. This is a subtle but significant issue in medical text generation, and the proposed ARW strategy provides a simple yet effective remedy. 2) RAM elegantly regularizes feature drift using a non-trainable anchor, with ablation results confirming its stabilizing effect on cross-modal alignment. 3) The two-stage training paradigm is clearly defined. The quantization of the consistency reward signal demonstrates careful engineering to ensure stable policy updates during GRPO. 4) The use of both radiologist evaluation and LLM-as-a-judge for reasoning quality (Table 2) compensates for the limitations of surface-level metrics (BLEU/ROUGE) in assessing clinical logic.
- 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) Figure 2(c) shows that 2D medical VLMs utilizing GRPO-based RL still underperform Med3D-R1 (S2). The paper attributes the gain to the 3D model and consistency reward but does not elaborate on why a similar RL strategy fails to yield comparable gains for 2D models. 2) Table 3 shows a sharp decline in CIDEr and BLEU when λ > 1.10, attributed to “overemphasizing abnormalities. “ A qualitative example would strengthen this observation. 3) Table 5 clearly indicates that the consistency reward’s efficacy is contingent on using RM-Mistral-7B, while Qwen/Llama provide suboptimal or harmful supervision.
- 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.
(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 paper presents a compelling and technically sound framework for 3D medical VLMs. The identification and mitigation of structural narrative bias is a significant and novel contribution to the field of medical report generation. The experimental validation is thorough, and the results convincingly demonstrate state-of-the-art performance.
- 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.
All my confusions about Fig 2, Table 3 and Table 5 have been explained clearly by the author. The paper could be accepted due to the deep insights on 3D medical VLMs for CT diagnosis.
Review #2
- Please describe the contribution of the paper
This paper presents Med3D-R1, a 3D medical vision-language framework for CT understanding that combines three components: a Residual Alignment Mechanism (RAM) to stabilize 3D-text alignment, an Abnormality Re-Weighting (ARW) strategy to mitigate positional bias in radiology reports, and a reasoning-aware RL stage with a consistency reward intended to improve the quality of intermediate reasoning traces. Overall, I think this is a solid and interesting paper. It is well written, the motivation is clear, and the experiments are relatively comprehensive. The main limitation is that the evidence for improved reasoning is still somewhat indirect.
- 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.
One main strength of this paper is that the motivation is quite compelling. The authors identify an interesting form of structural narrative bias in CT reports: normal findings tend to appear earlier while abnormal findings are concentrated later in the report. They further argue that this may bias autoregressive models toward over-predicting normal findings. I think this is a meaningful observation, and it gives the work a clearer clinical motivation. Another strength is that the method is easy to understand. Each component is designed to solve a specific problem: RAM addresses unstable cross-modal alignment, ARW targets report bias, and the consistency reward in the RL stage is used to reduce reasoning hallucination. Because of this, the whole framework feels coherent. The experiments are fairly sufficient. The paper includes comparison with multiple baselines, ablation studies on the proposed modules, and additional analysis such as reward quantization and reward model comparison. These results provide reasonabe support for the effectiveness of the method.
- 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 biggest issue, in my opinion, is that the claim about improving reasoning quality is still not fully convincing. The proposed consistency reward mainly encourages the generated reasoning trace to be more similar to the reference report, but this does not necessarily mean the reasoning itself is truly correct or clinically faithful. In other words, the method may improve the consistency of generated text more than the actual validity of the reasoning process.
- 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?
This paper presents Med3D-R1, a 3D medical vision-language framework for CT understanding that combines three components: a Residual Alignment Mechanism (RAM) to stabilize 3D-text alignment, an Abnormality Re-Weighting (ARW) strategy to mitigate positional bias in radiology reports, and a reasoning-aware RL stage with a consistency reward intended to improve the quality of intermediate reasoning traces. Overall, I think this is a solid and interesting paper. It is well written, the motivation is clear, and the experiments are relatively comprehensive. The main limitation is that the evidence for improved reasoning is still somewhat indirect.
- 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 paper addresses the development of 3D VLMs with robust clinical reasoning. It describes three challenges in doing so, i.e., (1) structural bias in radiology reports, (2) stabilizing cross-model alignment while aligning 3D vision features with textual features, and (3) reducing hallucinations in intermediate reasoning chains. The authors propose an abnormality re-weighting strategy using DeepSeek-V3 for sentence segmentation to address (1). A residual alignment module is proposed to address the challenge described in (2); lastly, a consistency reward is proposed for (3). The paper addresses an interesting set of challenges relevant to the medical MLLM/VLM literature.
- 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 three challenges identified in the paper are quite interesting and are relevant to the current VLM landscape of problems. The individual contributions made to address the three identified challenges are supported by ablation studies. The paper depicts an improvement in performance on CT-RATE and RAD-ChestCT for abnormality diagnosis over existing 3D VLMs.
- 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.
a. While the three contributions made in the paper are quite interesting, the paper lacks strong evidence to motivate them. For example. Fig 1.shows the structure followed in the reports. While this is an interesting observation, I’m curious to know if such a structure is followed only in CT-RATE/RAD-ChestCT or all types of radiological reporting done globally? Additionally, is there any merit in adhering to such a structure for a radiologist (but is possibly detrimental to the model)? Highlighting this aspect will motivate this challenge better to show how emerging techniques need to be modified in order to adhere to standard clinical practices. The current manuscript seems to have a disconnect with the clinical nuance for this statement. b. In section 2.1, while the RAM seems like an interesting approach, the conclusion of RAM reducing feature drift and yielding more stable 3D text alignment is not backed by an intuitive explanation on why it would help to do so. While the results depict an improvement, the methodology section needs to hint at strong reasoning on why the proposed method seems viable over other existing alignment techniques. Without this reasoning, it’s hard to understand how the proposed solution seems to be providing superior conceptual novelty over existing methods. c. In section 2.2, the usage of DeepSeek or any other LLM needs to be clearly motivated. If the structure of the reports is fairly standard (Fig 1.), then one can also apply a non-LLM-based technique (such as using regex, etc.) here, as this would not allow the method to get computationally expensive. The reasoning behind why the sentence segmentation step needs to be carried out by another LLM needs to be justified, even briefly so, in the paper. d. In section 2.3, the authors describe that existing RL-based medical VLMs such as those in [13, 17 from the paper] optimize only the final answer, making the intermediate reasoning clinically implausible. Once again, this argument is not supported by strong evidence, such as by showing a few examples of intermediate reasoning chains from [13,17]. The lack of evidence in support of this observation makes it less appealing. e. Continuing on the previous point, the paper does not describe why supervising only the final output corrupts the intermediate reasoning in the first place or cite any existing literature examining this phenomenon. The misalignment between visual and textual features could be one of the possible reasons, but as the method already uses an alignment mechanism, I’m curious to know what could be the possible reason behind this effect. Examining this aspect would make the observation about intermediate reasoning much stronger and thereby support the novelty of the solution. f. I would recommend the authors to cite the works used for comparison in the main results i.e Fig 2.c and the ablations tables, to relate to the main text. g. While Table 2.shows an ablation for Consistency reward, a one-on-one comparison with [13, 17] is needed to show the improvement in intermediate reasoning chains after applying the proposed method, along with the overall improvement. h. In section 3.2, the language model that is used as the decoder needs to be specified. i. As the manuscript does not provide any information of a source code release, it requires detailed specifications of hyper-parameters for reproducibility.
- 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.
(2) Reject — should be rejected, independent of rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
While this paper does talk about some interesting aspects in the development of Medical VLMs but the insufficient evidence supporting the arguments and observations make the paper less appealing. The organisation of the paper also needs improvement so that main results and the effect of the individual contributions are highlighted properly.
- 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 authors for their rebuttal. The authors provide a reasonable rebuttal, and they resolve some of my concerns. The clarification provided in the rebuttal is essential and would make the draft stronger. As the hyper-parameter details are quite essential for such setups, a code release becomes crucial. The authors have committed to the same, if accepted. I would also encourage them to incorporate the missing details as pointed out in the review in the final/updated version.
Author Feedback
We thank all reviewers and ACs for their valuable feedback. We are encouraged that R1 and R2 recognized the novelty and empirical strength of Med3D-R1, and following the meta-review, we focus on the major concerns regarding clinical motivation, technical justification, and evidence of reasoning improvement. 1.Clinical motivation of structural report bias (R3-a, Meta) R3 questions whether the normal-before-abnormal structure is dataset-specific. We clarify that our empirical claim is limited to the submitted training corpus, where Fig. 1 quantifies a clear positional regularity. More broadly, this pattern is consistent with structured radiology reporting practice, where normal anatomy is often documented systematically before abnormal findings. Similar tendencies are also observable in structured report datasets such as MR-RATE, while our formal quantitative claim remains limited to the submitted corpus. Our point is not to criticize this clinically useful convention, but to show that autoregressive training can turn it into a harmful positional prior. 2.Technical justification of RAM and DeepSeek-based sentence labeling (R3-b/c) RAM addresses the instability of mapping high-variance 3D tokens into the lower-variance language space. The fixed anchor is not semantic evidence, but a stable center of the pretrained text-embedding space. Thus, the projector learns bounded residual deviations rather than an unconstrained full mapping, reducing extreme tokens and feature drift. Table 1 supports this: removing RAM decreases all report metrics and especially semantic scores (BERT_F1/LLM_Score). For DeepSeek-V3, we clarify that it is not used for sentence segmentation. Reports are first split into sentences by rule-based preprocessing. DeepSeek-V3 is only used to label each sentence as normal, abnormal, or uncertain. Regex can split sentences or sections, but is less reliable for negation, uncertainty, and nuanced radiological descriptions. This offline step is used only for training-data construction and adds no inference overhead. 3.Evidence-consistent reasoning improvement (R2, R3-d/e, Meta) R2/R3 note that consistency with reference reports does not guarantee clinically faithful reasoning. We agree and clarify the scope: Med3D-R1 does not claim absolute clinical truth, but aims to improve evidence-consistent reasoning under report-based supervision, encouraging reasoning traces to align with report-supported facts and penalizing unsupported intermediate statements. Thus, Table 2 includes radiologist assessment alongside LLM-as-a-judge evaluation, rather than BLEU/ROUGE. These results support improved clinical faithfulness within the scope of available report supervision. 4.Comparison with prior RL-based VLMs (R3-g) R3 requests a direct reasoning-chain comparison with [13,17]. Fig. 2(c) already provides a one-on-one MVQA comparison with Med-R1 and MedVLM-R1, where Med3D-R1 achieves higher diagnostic accuracy. In addition, Table 2 isolates the effect of the consistency reward within our framework, and radiologist evaluation supports improved clinical faithfulness. These results suggest that evidence-consistency supervision improves reasoning reliability beyond final-answer rewards. 5.Source of gains and reward design (R1) Fig. 2(c) does not attribute gains to RL alone: S1 gives 3D evidence and RAM-stabilized alignment; S2 adds consistency reward. The λ>1.10 drop reflects reduced textual fidelity from abnormal overemphasis. Table 5 suggests discriminative RM-Mistral is more reliable than generative judges for reward scoring. 6.Minor clarifications and reproducibility (R3-f/h/i) We thank R3 for pointing out these missing details. The language decoder is Qwen2.5-3B-Instruct, which will be explicitly specified in the final version. We will also add citations for the compared methods used in Fig. 2(c) and the ablation tables. We are committed to releasing the code with implementation details and the trained model upon acceptance.
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.
Given the strong positive feedback from two reviewers on the method’s novelty and performance, while significant unresolved questions from the third reviewer regarding clinical motivation and technical justification, authors are intived to rebuttal. The authors should address the key points towards the clinical motivation and evidence of reasoning improvement.
- 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.
After rebuttal, all reviewers would like to accept the submission.
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.
This paper presents Med3D-R1 for CT-based diagnosis. Although the initial reviews were mixed (Accept/Weak Accept/Reject), all reviewers are happy with the authors’ rebuttal and lean to acceptance. The AC concurs with the reviewers’ updated assessments. The additional discussions in the rebuttal should be incorporated into the final version.
Meta-review #3
- 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 provided sufficient details in the rebuttal which addressed R3’s comments. I recommend Accept considering that it is likely the first reasoning 3D VLM and would raise discussion at MICCAI 2026.
