List of Papers Browse by Subject Areas Author List
Abstract
Medical Vision-Language Models have shown promising potential in clinical decision support, yet they remain prone to factual hallucinations due to insufficient grounding in localized pathological evidence. Existing medical alignment methods primarily operate at the response level through preference optimization, improving output correctness but leaving intermediate reasoning weakly connected to visual regions. Although chain-of-thought (CoT) enhances multimodal reasoning, it remains largely text-centric, limiting effective integration of clinical visual cues. To address this gap, we propose ClinCoT, a clinical-aware visual chain-of-thought framework that transforms preference optimization from response-level correction to visual-driven reasoning. We introduce an automatic data generation pipeline that constructs clinically grounded preference pairs through reasoning with hypotheses-driven region proposals. Multiple Med-LLMs evaluators rank and assign scores to each response, and these rankings serve as supervision to train the target model. We further introduce a scoring-based margin-aware optimization strategy that incorporates both preference ranking and score difference to refine region-level reasoning trajectories. To maintain alignment as the model’s policy evolves during training, we adopt an iterative learning scheme that dynamically regenerates preference data. Extensive experiments on three medical VQA and report generation benchmarks demonstrate that ClinCoT consistently improves factual grounding and achieves superior performance compared with existing preference-based alignment methods.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/3633_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to the Code Repository
https://github.com/Xiwei-web/ClinCoT
Link to the Dataset(s)
N/A
BibTex
@InProceedings{LiuXiw_ClinCoT_MICCAI2026,
author = { Liu, Xiwei AND Li, Yulong AND Zhuang, Xinlin AND Li, Xuhui AND Chen, Jianxu AND Yang, Haolin AND Razzak, Imran AND Xie, Yutong},
title = { { ClinCoT: Clinical-Aware Visual Chain-of-Thought for 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
The authors propose ClinCoT, a clinical-aware visual chain-of-thought (CoT) framework designed to improve factual grounding in Medical Vision-Language Models (Med-VLMs) by hypothesis-driven, region-level reasoning, like how case-reviews can be conducted. The framework introduces an automatic two-stage pipeline: (1) disease-conditioned region proposals generated via a clinical-aware tool, and (2) consensus-weighted quality assessment using two Med-LLM evaluators (LLaMA3-Med42-7B and BioMistral-7B) to construct preference pairs. Using LLaVA-Med v1.5 as the target model, ClinCoT is evaluated on two medical VQA benchmarks (SLAKE, VQA-RAD) and one report generation benchmark (IU-Xray). ClinCoT achieves incremental improvement in performance on report generation and competitive results on VQA tasks, particularly under supervised fine-tuning (SFT) settings.
- 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.ClinCoT is well clinically motivated. The hypothesis-driven region proposal mechanism is mimics how radiologists’ reason and form potential differential diagnoses. By examining localized regions rather than processing entire images uniformly, ClinCoT can generate multiple hypothesis driven answers, giving the framework strong conceptual grounding. 2.Consensus weighted scoring (Eq. 4) using two independent Med-LLM evaluators allows for the penalization of disagreement, while leveraging pair construction allows for an optimized forward reasoning trajectory that grounds the model in its preferred track. Critically, the explicit construction of both a dispreferred and a preferred reasoning chain forces the model to learn discriminative boundaries between clinically valid and invalid reasoning trajectories. This dual-chain design anchors the model’s intermediate reasoning steps to localized pathological evidence, producing a more grounded and stable reasoning trajectory particularly beneficial for VQA answer generation. 3.ClinCoT leverages iterative learning to address distributional shift. The iterative preference data regeneration scheme is a principled adaptation of online DPO to the medical domain, preventing stale preference data from degrading alignment as the model’s policy evolves.
- 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.Comparing ClinCoT and MMedPO, we see mixed VQA. ClinCoT underperforms MMedPO on VQA-RAD open-ended recall (33.15 vs. 36.36 without SFT), which the authors attribute to instability in intermediate reasoning for short-form answers. However, no concrete ablation or analysis is provided to validate this hypothesis or to clarify under what conditions region-conditioned CoT is reliably beneficial or fails. 2.The computational cost is not discussed in the manuscript. The pipeline involves multiple Med-LLM evaluator calls per training sample across m=4 iterative rounds, which is likely computationally expensive. No wall-clock time, compute cost, or efficiency comparison against baselines is provided. Depending on clinical inference time, this approach would not be ideal for live-patient / radiologist interaction but could be optimized for overnight data review and patient summaries for day-staffs.
- Please rate the clarity and organization of this paper
Good
- Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.
The submission has provided an anonymized link to the source code, dataset, or any other dependencies.
- Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?
N/A
- Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2026/en/REVIEWER-GUIDELINES.html
The paper addresses a genuine and important gap in existing medical VLM alignment methods, improving VQA outputs without explicitly reasoning over localized pathological evidence. The SFT+ClinCoT results in Table 1 are particularly interesting and suggests the method works best as a refinement layer on top of a domain-adapted base. Given the results, it would be interesting for future work to examine robustness to noisy region and artifacts, and expend evaluation to include more non-chest-X-ray modalities.
- 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?
ClinCoT presents a well-motivated and technically sound framework with a comprehensive ablation study and competitive empirical results. The core idea of embedding region-level clinical reasoning into preference optimization is novel and clinically relevant. However, the inconsistent VQA performance compared to MMedPO, and the unexplored sensitivity to region proposal quality leave open questions about generalizability. A weak accept is appropriate given those limitations.
- 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
The authors present ClinCoT, a clinical-aware visual chain-of-thought (CoT) framework that shifts preference optimization from response-level correction to visual-driven reasoning. ClinCoT aligns visual grounding with clinical inference through intermediate reasoning trajectories, employing two approaches: hypotheses-driven region generation and consensus-weighted quality assessment. The hypotheses-driven region generation produces clinically grounded preference pairs by using a clinical-aware tool to identify disease-conditioned regions, while the consensus-weighted quality assessment scores these pairs using multiple medical large language model (LLM) evaluators. The model is then fine-tuned with a margin-aware direct preference optimization (DPO) and an iterative preference learning strategy.
- 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 proposed study addresses an important limitation of current Medical Vision-Language Models (Med-VLMs). Many existing methods remain predominantly text-centric and do not explicitly model how localized visual evidence shapes intermediate reasoning. In contrast, the proposed hypotheses-driven visual reasoning framework is clinically well-motivated.
2.The proposed margin-aware DPO leverages differences in preference scores, which may more accurately capture the unequal contributions of various regions to the final decision for finer discrimination between preference pairs.
3.The proposed automatic data generation pipeline is scalable and clinically meaningful, as it leverages intermediate reasoning conditioned on localized contexts derived from the clinically grounded preference pairs.
4.The authors evaluate ClinCoT on three benchmarks: two medical visual-question answering (VQA) datasets (VQA-RAD and SLAKE) and one report generation dataset (IU-Xray). They compared ClinCoT against seven baselines, including standard DPO and its variants: DPO, self-rewarding, STLLaVA-Med, POVID, SIMA, FiSAO, and MMedPO.
- 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 authors note that most existing CoT approaches are text-centric. However, the study does not include a direct performance comparison with a text-centric CoT baseline.
2.In Section 2.1, the authors mention that disease-conditioned region hypotheses are obtained using a clinically aware visual tool. This suggests that the proposed method heavily depends on the quality of the disease-conditioned activation maps. However, the study does not validate the anatomical or clinical accuracy of the extracted regions.
3.In Section 2.1, two Med-LLM evaluators are employed in the consensus-weighted scoring strategy to assess the quality of all generated responses. However, using two Med-LLM evaluators may raise practical concerns, including high computational costs, reproducibility, and potential evaluator bias. Additionally, the study does not address how closely these evaluators’ performances correlate with those of medical professionals or experts.
4.Including a qualitative review of the reasoning chains by radiologists to evaluate whether the generated chains follow clinically appropriate inference paths would strengthen the study.
5.In Section 3.2, the results indicate that ClinCoT underperforms compared to MMedPO on the VQA-RAD dataset in the non-supervised fine-tuning (SFT) setting. This suggests that the advantages of CoT-based reasoning may depend on the specific task characteristics and training configurations. It would be valuable to discuss this observation more explicitly.
- 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
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?
Several key issues prevent the paper from being accepted at this stage. First, the paper does not include a direct performance comparison with text-centric CoT baselines, which weakens its major claim of improving upon prior reasoning paradigms. Second, the framework heavily depends on the quality of the disease-conditioned region hypotheses, but the study does not validate the anatomical or clinical accuracy of these regions, raising concerns about the reliability of the visual grounding. Third, the use of two Med-LLM evaluators in the consensus-weighted scoring strategy raises practical concerns regarding computational cost, reproducibility, and potential evaluator bias, which are not sufficiently addressed. Finally, the underperformance on VQA-RAD dataset compared to MMedPO suggests that the benefits of the proposed approach may not generalize consistently across tasks, and this inconsistency is not adequately analyzed.
Overall, while the paper presents promising ideas and has the potential for significant impact, the current empirical validations and analyses are insufficient to fully support its claims. I would consider a higher score if these concerns are addressed in the rebuttal.
- 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
This paper proposes a clinically perceived visual thinking chain alignment framework (ClinCoT): 1.Data generation: An automated pipeline is introduced to construct a preference pairing of “visual alignments” (Preferred vs. Dispreferred) based on lesion hypothesis-driven regional proposals. 2.Alignment strategy: A consensus-weighted marginal perception DPO optimization method is designed, which takes into account not only the ranking but also the score differences. 3.Iterative learning: A dynamic update mechanism is employed to mitigate distribution shifts caused by the evolution of model strategies during the training process.
- Please list the major strengths of the paper: you should highlight a novel formulation, an original way to use data, demonstration of clinical feasibility, a novel application, a particularly strong evaluation, or anything else that is a strong aspect of this work. Please provide details, for instance, if a method is novel, explain what aspect is novel and why this is interesting.
It highlights the limitations of existing CoT approaches, which place too much emphasis on text and neglect visual cues. This direction aligns well with the actual logic of radiological diagnosis. By utilizing multiple medical large language evaluators and calculating the degree of agreement to penalize controversial assessments, the quality of the automatically generated supervised signals is improved.
- 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 quality of regional proposals in the framework highly depends on pre-trained visual tools such as MedKLIP. If the front-end tools fail to detect certain issues, the subsequent inference processes will also deviate from the correct path. This cascading error effect is not fully discussed in the text. Using Med-LLM as a “judge” to train Med-VLM may result in the model merely adopting the biases of the evaluator.
- 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
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 main advantage of this study lies in the transformation of the approach to preference optimization from a simple response-level correction to a process-driven visual reasoning paradigm. This addresses the issue where the reasoning processes of existing medical visual language models are often disconnected from specific pathological evidence. Furthermore, the model employs an automated, hypothesis-driven method to generate regional-level preference data, providing a scalable solution to the problem of medical image alignment without the need for expensive manual annotation by experts.
- 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.
Reviewers 1 and 3, recommend acceptance, praising the clinical motivation and novel visual CoT framework, and Reviewer 2 recommends Weak Reject due to the lack of comparison with text-centric CoT baselines, unverified accuracy of region proposals, and inconsistent performance on VQA-RAD. Overall, the merits overweight the weakness of this paper, therefore, I tend to accept this submission.