List of Papers Browse by Subject Areas Author List
Abstract
Large vision-language models (LVLMs) are increasingly used for clinical image understanding, yet they remain vulnerable to \emph{hallucinations}–producing textual findings or attributes not supported by the image. We present a vision-traceable hallucination detection framework that audits arbitrary LVLM responses via visual evidence grounding, requiring neither modification nor internal access to the hidden states of LVLMs. Given an LVLM response, we extract visually verifiable entities and use a medical-domain-adapted Qwen-VL grounding verifier to localize each entity on the input image. To enhance the robustness of our detection method, we introduce a counterfactual entity perturbation method and estimate visual evidence uncertainty by contrasting factual and counterfactual grounding results. Specifically, we compute an entity-level uncertainty score from the positive confidence, counterfactual confidence, and their grounding overlap for binary hallucination decision-making. Experiments on multiple medical imaging modalities and LVLM backbones demonstrate that our method consistently improves hallucination detection performance over recent baselines, while providing interpretable localization evidence and strong cross-model transferability. Code and dataset are available at \url{https://github.com/Agentic-CliniAI/CounterVHD}.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/4284_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to the Code Repository
https://github.com/Agentic-CliniAI/CounterVHD
Link to the Dataset(s)
https://github.com/Agentic-CliniAI/CounterVHD
BibTex
@InProceedings{SonXia_Detecting_MICCAI2026,
author = { Song, Xiao AND Qin, Haonan AND Zhang, Zhaoxu AND Zhang, Jiong AND Fang, Yuqi AND Shan, Caifeng},
title = { { Detecting Clinical Hallucinations in LVLMs via Counterfactual Visual Grounding Uncertainty } },
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
The paper proposes a plug-and-play hallucination detection method for medical LVLMs that does not require access to the underlying model’s internal states. Visually verifiable entities are extracted from an LVLM response (via GPT-4.1-mini) and grounded on the input image using a medical-domain fine-tuned Qwen3-VL-2B verifier, yielding a box and confidence se+s^+_e se+. For each entity, a mutually-exclusive counterfactual (e. g. “right kidney” → “left kidney”) is also grounded, yielding (be−,se−)(b^-_e, s^-_e) (be−,se−). An uncertainty score Ue=1−(se+−IoUe⋅se−)U_e = 1 - (s^+_e - \text{IoU}_e \cdot s^-_e) Ue=1−(se+−IoUe⋅se−) is thresholded for a binary hallucination decision. The method is evaluated on a CT/MRI test set of 1,904 image–report pairs built from IMIS-Bench with hallucinations induced via counterfactual label substitution using four LVLMs (MedGemma-27B, GPT-5.1, Gemini-3-Flash, Grok-4-Fast), against three general-domain 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.
**Clear, intuitive idea. ** Grounding a contralateral/mutually-exclusive counterpart as a consistency check is a simple and compelling framing.
**Model-agnostic by construction. ** The verifier is decoupled from the audited LVLM, so the method is applicable to closed-source models.
**Interpretability. ** Returns bounding-box evidence alongside a decision, which is clinically useful.
**Strong MRI results. ** On MRI, the method dominates baselines on both HDR and precision across all four LVLMs.
**Ablation supports the CF branch. ** MCS AUC drops from 0.88 to 0.36 without CF; IoU reweighting gives an additional small gain (AUC 0.86 → 0.88 for MCS).
**Efficiency. ** A 2B verifier is competitive against a 13B FaithScore pipeline.
- 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.
**Prior Art could be disciussed more. ** [2503.20504] UniVRSE: Unified Vision-conditioned Response Semantic Entropy for Hallucination Detection in Medical Vision-Language Models is cited, but at a bit strange point, and not discussed, even though this method could also be described as similarly counterfactual (distorting image and evaluating prediction changes).
**Construct validity: test set and method share the same inductive bias. ** Test hallucinations are “induced via mutually exclusive counterfactual labels” — the same substitution principle the CF branch is designed to detect. Established benchmarks (POPE, Med-HallMark, MedHEval) construct hallucinations via mechanisms independent of any specific detector. The present construction likely inflates results, particularly on MRI where laterality is salient, and does not test detection of other hallucination categories (fabrication, omission, severity, measurement, confabulated priors).
**MCS-only AUC of 0.3572 is below-random and diagnostic. ** Below-random AUC typically signals score-direction inversion or anti-correlation between score and label, not a “weak” detector. Under sign flip, AUC = 0.6428 (a reasonable standalone signal). The authors should report the sign-flipped AUC and per-category breakdowns; as presented, the ablation may be overstated.
**Incomplete metrics. ** Only HDR (recall) and precision are reported in Tab. 1.POPE, MedHallBench, etc. all report F1, which makes it easier to compare approaches along a single metric.
**Writing. ** Abstract: “Code and datasets will be accessed. “ Section 3.2 contains a fragment: “our visual-aware ver and the counterfactual-driven Therefore…”. Fig. 4 sub-panel labels (a)/(b) appear inconsistent with the caption and Sec. 3.4 text.
- 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
The formula for the uncertainty seems a bit ad-hoc, have you evaluated any alterantive formulas?
- 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?
It is an interesting core idea, some more related work could be discussed and further metrics reported.
- 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
This paper introduces a method to more reliably trace or detect hallucinations in VLMs for medical imaging tasks. The proposed solution consists of two main components, i) a verifier model (Qwen-VL fine-tuned on medical domain tasks) to locate the VLM-predicted response entities in the input image, and ii) counterfactual generation for uncertainty quantification.
- 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 proposed approach seeks to explicitly localize the predicted entities in the image, which adds to the reliability and robustness of the method.
The illustrated pipeline in Figure 2 gives a clear overview of the method, which I find to be generally well-motivated 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.
More details should be included on how the verifier model is trained / fine-tuned, since this is critical to the success of the proposed approach. Specifically, I am also interested in how domain gaps between the verifier model and the medical imaging task could possibly affect the results. It would be more convincing if the verifier can generalize to different datasets (beyond the IMIS-Bench used in this study). Otherwise, there is not much value in the verifier approach should we want to extend to real world applications where there could be large domain differences.
The description concerning counterfactual generation is lacking in details. It is totally unclear how the different components in equation 4 come about and again, the paper would benefit from clarity for this portion. Other related questions that I have is how do the authors ensure the non-overlapping constraint. How should the counterfactual generation proceed if this non-overlapping constraint cannot be satisfied?
- 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?
This paper proposed a well-motivated reasonable approach to ground VLM diagnosis visually and detect possible hallucinations in a more robust fashion. However, some details are missing for key components of the approach, as mentioned in the weakness section. I would encourage the authors to include complete descriptions of these details and I will raise my score if my questions are clarified.
- 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 authors have included more details regarding the training of the verifier model and counterfactual generation. This has clarified my main concerns and I would raise my score to acceptance. I would also recommend that these details be included in the final revision as they would be helpful for the readers.
Review #3
- Please describe the contribution of the paper
The authors present an interesting approach for addressing clinical hallucinations using LVLMs that works on multiple modalities and provides spatial information. The contribution relies on a novel alternative for hallucination detection based on counterfactual-driven uncertainty estimation, which identifies false lesion evidence by contrasting factual and counterfactual grounding. The main advantage of this approach is that the authors not only provide an above the state-of-the-art performance, but also add value to the method by enabling traceability and interpretability of the results. All of this presents a clear advantage for clinicians and future research on the matter.
- 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.
This paper proposes a novel method for identifying clinical hallucinations via counterfactual entity perturbation, which has the potential to inspire further research in this area and to support future translation into clinical practice.The presented methodology uses the alignment of responses with visual regions, which improves the transferability to the clinical setting by enhancing interpretability. In addition, this novel methodological research approach opens the door for new ways of tackling misdiagnosis and therefore more robust and reliable AI in clinical practice, as results have been shown to be above state-of-the-art methodologies. The main highlight is the plug-and-play and vision-traceable methodology that also enables uncertainty estimation by means of counterfactual entity perturbation, which is a very interesting leveraging of a technique for probing and improving model behavior that has proven relevant in sensitive domains like clinical AI/ML.
- 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.
Counterfactual entity perturbation is a useful approach for detecting hallucinations, but it also comes with several limitations. One of the main concerns is the realism of the generated perturbations, which may not always correspond to clinically plausible variations, especially in complex imaging scenarios where maintaining anatomical consistency is challenging. As a result, some counterfactual examples may fall outside the data distribution, making the model’s responses difficult to interpret. In addition, it is not clear whether the method establishes true causal relationships, as it can overlook dependencies between features and thus simplify underlying clinical factors. Its effectiveness can also depend heavily on how the perturbations are designed, which requires careful consideration and domain expertise. Furthermore, generating and evaluating counterfactuals can be computationally intensive, and there is a risk of interpreting changes in model predictions as clinically meaningful when they may instead reflect artifacts of the model. Including a dedicated discussion of these limitations would help provide a more balanced view of the approach and encourage further work in this area.
- Please rate the clarity and organization of this paper
Good
- Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.
The authors claimed to release the source code and/or dataset upon acceptance of the submission.
- Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?
N/A
- Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2026/en/REVIEWER-GUIDELINES.html
Figures would benefit from a font size increase for readability. This sentence appears incomplete: “We attribute this gain to our visual-aware ver and the counterfactual-driven Therefore, the consistent performance…”
- 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.
(6) Strong Accept — must be accepted due to excellence
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
This paper presents an interesting approach for a widely encountered problematic in the clinical setting when using machine learning algorithms for lesion detection, as hallucinations can lead to mis-diagnosis. It leverages a novel technique for detecting clinical hallucinations using counterfactual entity perturbations with a plug-and-play, interpretable and vision-traceable method, all of which can have an impact by opening the door to future research in this area while exhibiting potential for its transferability to the clinic.
- Reviewer confidence
Very confident (4)
- [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.
Authors addressed all my comments and suggestions, my recommendation in the prior stage was already acceptance.
Author Feedback
We sincerely thank reviewers for constructive feedback: R1 finds the core idea “interesting, clear and intuitive”, and praises its “clinical interpretability” and “efficiency (2B vs. 13B)”; R2 considers it a “well-motivated reasonable approach” and will raise the score if more details are clarified; R3 rates Strong Accept (6/6): “…a plug-and-play and vision-traceable method…can have an impact by opening the door to future research in this area while exhibiting potential for its transferability to the clinic.” Responses below.
1.UniVRSE could be discussed more (R1Q1) 1) Thanks for highlighting the UniVRSE arXiv preprint. Our citation [Ref. 9] specifically supports taxonomy in Sec. 1 and Fig. 1, which was informed by its peer-reviewed MICCAI 2025 version available at our submission time. So, the citation is correct. 2) In the revised version, we will add a discussion of UniVRSE. It introduces an image-distortion-based semantic uncertainty algorithm (aligns with Fig. 1(b)) that is effective for Chest X-ray report generation and VQA. In contrast, we provide direct visual evidence for interpretability (Fig. 1(c)).
2.Whether the test set shares the same inductive bias as the detector (R1Q2) 1) Test set comes from a static perturbation across four model families (Sec. 3.1), while the detector applies counterfactual reasoning for dynamic consistency verification, independent of our method. Thus, no shared inductive bias exists. 2) Benchmarks like POPE (noted by R1) evaluate LVLM hallucinations while we evaluate detection methods. Given this fundamental difference, their direct relevance to our work is limited.
3.Other hallucination types (R1Q2) To maintain focus, this work concentrates on “text-image consistency” (see Abs). R1 noted other categories (e.g., severity, omission), but they varied fundamentally in detection logic (e.g. omitting regular and normal findings is clinically routine rather than hallucinatory), warranting dedicated future study.
4.Missing F1 and issue about MCS-only AUC (R1Q3, R1Q4) 1) F1: We can directly calculate F1 from Recall and Precision in Tab. 1.2) AUC: We thank the reviewer for the score-direction insight. MCS-only (w/o CF), sign-flipped AUC=0.6428, is a reasonable signal; with CF, AUC rises to 0.8811 (+0.2383), confirming CF’s necessity (Fig. 3(b)).
5.Details on verifier training and domain-gap robustness (R2Q1) 1) Training: LoRA fine-tuning on Qwen3-VL-2B for visual grounding, AdamW (lr=1e-4, wd=1e-4), CosineLR (3% warmup), bs=16 with 8-step grad accumulation, 5 epochs, 2×A100, 33,266 samples from IMIS-Bench. 2) Domain-gap: Our core contribution is designing a plug-and-play method applicable to arbitrary LVLMs (Sec. 2.1), where addressing the domain gap is beyond our main focus. However, we still mitigate it by training the verifier on IMIS-Bench (Sec. 3.1), a collection of 80 datasets. This broad coverage largely reduces the domain gap.
6.Concerns about counterfactual details and Eq.4 (R2Q2) Eq. 4 retrieves a counterfactual perturbation of e (from Eq. 1-2) from knowledge base K. The non-overlap guarantee is built into K’s construction: we apply a strict spatial exclusion rule that discards any entity whose bounding box could overlap with others on identical MRI/CT slices, which was independently verified by radiologist to ensure clinical validity.
7.Causality and cost (R3) 1) Causality and future directions: Thanks for highlighting these important considerations regarding realism and causal validity. We fully agree that addressing confounding effects and modeling complex clinical dependencies are critical. Encouragingly, some directions align with the future work outlined in our Conclusion, specifically the extension to complex spatial relationships and temporal sequences. 2) Cost: Our 2B verifier is more efficient than Faithscore (13B)(Tab. 1), so the consumption of a CF branch is acceptable.
8.Writing (R1Q5) We thank the reviewer. All will be corrected in the revised 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.
The reviewers request rebuttal on the construct validity of the benchmark, since the hallucination test set appears to share the same counterfactual substitution bias as the proposed detector, which may inflate performance and limit conclusions beyond label-conflict hallucinations.
The authors should clarify key methodological details, especially verifier training/fine-tuning, domain-gap robustness, counterfactual generation, non-overlap constraints, and what happens when clinically plausible counterfactuals cannot be generated.
The rebuttal should also address missing comparisons and metrics, discuss related counterfactual/hallucination methods such as UniVRSE more clearly, report fuller evaluation results including F1 and sign-flipped/per-category AUC for the MCS ablation, and add a balanced discussion of counterfactual realism, causality limits, and computational cost.
- 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 paper is accepted. All three reviewers are positive after rebuttal, with two reviewers explicitly confirming that their main concerns were addressed and another already supporting acceptance. The work tackles an important clinical problem—hallucination detection in medical VLMs—and proposes an interpretable, plug-and-play, counterfactual entity perturbation framework with clear potential to stimulate future research. The final version should include the added verifier-training and counterfactual-generation details for clarity and reproducibility.
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.
R2 changed their score to Accept and recommended certain details to be added to the camera-ready 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.
Reviewers have followed up acknowledging that most raised concerns were addressed. A consensus has now been reached for acceptance. The reviewers have acknowledged the relevance of the work (locating hallucinations) and the well-justified advantages of the proposed approach.
