List of Papers Browse by Subject Areas Author List
Abstract
Transcatheter Aortic Valve Replacement (TAVR) planning requires meticulous multimodal reasoning. However, adapting Multimodal Large Language Models (MLLMs) to this high-stakes domain is severely impeded by diagnostic hallucinations, where generated text lacks anatomical grounding. To address this, TAVR-VLM is introduced: a novel framework featuring Risk-Conditioned Causal Grounding Attention (R-CGA) that instantiates a model-internal ``Risk $\rightarrow$ Region $\rightarrow$ Word’’ structural grounding pathway. R-CGA compresses multimodal inputs into a causal risk bottleneck, purifying dense visual features into a global risk mask. During autoregressive generation, a support-projected causal consistency objective constrains token-level grounding within the risk-defined support mask. Evaluated on $\mathrm{M}^3\mathrm{TAVR}$, TAVR-VLM demonstrates consistently strong performance across risk stratification, report generation, hallucination reduction and spatial grounding. The results indicate that R-CGA can enhance both the factual reliability and anatomical interpretability of vision-language models for TAVR preoperative planning.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/2099_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to the Code Repository
https://github.com/Leo1998-Lu/TAVR-VLM
Link to the Dataset(s)
N/A
BibTex
@InProceedings{LuZhi_TAVRVLM_MICCAI2026,
author = { Lu, Zhixiang AND Liu, Xiwei AND Song, Sifan AND Ji, Changkai AND Nguyen, Anh AND Su, Jionglong AND Razzak, Imran AND Wang, Jinfeng},
title = { { TAVR-VLM: Risk-Conditioned Causal Grounding for Hallucination-Resistant Report Generation } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 16878},
month = {September},
page = {pending}
}
Reviews
Review #1
- Please describe the contribution of the paper
The authors presents TAVR-VLM, a multimodal framework for Transcatheter Aortic Valve Replacement (TAVR) preoperative planning. The core innovation is the Risk-Conditioned Causal Grounding Attention (R-CGA) module, which enforces a top-down “Risk → Region → Word” causal hierarchy to mitigate diagnostic hallucinations in multimodal large language models (MLLMs). The authors curate a new multi-modal TAVR dataset (M³TAVR, 1,482 patients) and demonstrate that their method outperforms specialized baselines and closed-source frontier models (e. g. , GPT-4o, Gemini-3 Pro) across risk prediction, report generation, and safety/grounding metrics. The framework achieves state-of-the-art results (AUROC 0.896, CIDEr 0.936, Hallucination Rate 8.1%, mIoU 0.624) via explicit causal constraints, validated by comprehensive ablation studies.
- 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 authors address a clinically critical and under-explored task: trustworthy preoperative report generation for TAVR with reduced diagnostic hallucinations. They propose a novel and structurally interpretable R-CGA framework that enforces a strict top-down causal hierarchy from risk hypothesis to visual grounding to report generation. Comprehensive benchmarking, thorough ablation studies, and parameter sensitivity analysis are conducted on a large-scale multi-modal TAVR dataset, showing consistent and significant performance gains over strong baselines including state-of-the-art open-source and closed-source VLMs. The framework effectively suppresses unsupported clinical claims via explicit causal constraints, offering verifiable, anatomically grounded diagnostic reasoning for high-stakes surgical decision-making.
- 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 reported high CIDEr score is potentially misleading and over-optimistic, as it is strongly inflated by the highly template-based TAVR report structure rather than genuine semantic and clinical superiority.
2.All experiments are limited to a single in-house dataset, and the method’s generalizability is not validated on public medical report generation datasets or external multi-center cohorts.
3.The definition and computation of the hallucination rate rely heavily on model-internal attention masks rather than independent human annotation, making it less objective and difficult to cross-validate.
4.The clinical translation potential is overclaimed. There is no evaluation on out-of-distribution or rare anatomical cases, and no expert user study to validate real-world clinical usability.
- 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?
The paper addresses a clinically relevant problem and proposes the R-CGA framework with effective performance gains, supported by comprehensive experiments. However, it has notable limitations (misleading CIDEr score, no external validation, less objective hallucination evaluation, overclaimed clinical translation). These issues limit a stronger acceptance but do not negate its value, leading to my recommendation of Weak Accept.
- 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 #2
- Please describe the contribution of the paper
1.The paper formalizes TAVR assessment as a risk-conditioned structured generation task.
2.The paper proposes a three-stage architecture comprising a Global Risk Bottleneck, Risk-Driven Visual Purification, and Token-Level Grounding to enforce a strict “Risk - Region - Word” reasoning hierarchy.
3.The paper implements a token-to-global mask alignment objective using a stop gradient mechanism.
- 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.
Risk assessment and recommendation for TAVR pre-intervention planning carries significant clinical value. The motivation to align the decision-making reasoning process with cardiologists’ established TAVR workflow is well-grounded and demonstrates clear clinical utility.
- 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 motivation behind this work is commendable; however, the central claim that “hallucination arises as a failure of the top-down causal link between global risk states and local anatomical evidence” requires stronger empirical or theoretical grounding. Given that this assertion serves as the primary design rationale for the proposed framework, the authors should provide more concrete supporting evidence to substantiate this claim before it can be fully accepted as a foundational premise.
2.The role and impact of the causal risk bottleneck warrants further clarification. As described in the paper, the risk score predicted in Stage 1 serves as the cornerstone of the global risk mask, risk-driven visual purification, and the final prediction. This makes the reliability of the risk estimator critical to the overall framework. The authors should provide a more thorough analysis of this component’s behaviour, in particular, what happens when the risk prediction is incorrect. A dedicated error analysis or robustness study examining the downstream effects of risk estimation failures would significantly strengthen the paper’s claims.
3.Several aspects of the methodology section remain insufficiently described. First, the distinctions between the discrete procedural risk levels ({l, m, h}) are not clearly defined. The authors should specify how these levels are operationalized and what criteria govern their boundaries. Second, the reconstruction loss L_rec appears undefined in the main text and should be formally introduced. Third, the process for selecting the subset of token indices corresponding to clinically salient entities when computing the Token-level Mask is unclear. Together, these gaps make the methodological section difficult to follow, and addressing them is essential for reproducibility.
4.The design of the risk estimation module raises two concerns. First, since the entire framework is trained from scratch, there is a risk of a chicken-and-egg problem: if the risk estimator is not yet well-trained in early stages, its errors may propagate and mislead the subsequent generation process, potentially causing compounding performance degradation. The authors should discuss how this instability is mitigated during training. Second, given the availability of numerous strong pretrained models for both CT and echocardiography feature encoding, the motivation for training the feature extractor from scratch is unclear. The authors should justify this choice and, ideally, compare it against fine-tuning pretrained encoders as a baseline.
5.Additional clarification is needed for parts of the experimental section. In the case study, Case A references a “mitral valve procedure” statement from the baseline VLM, but it is not made clear what specifically is problematic about this output. A brief explanation would make the qualitative comparison more informative. Additionally, the authors should indicate whether the proposed benchmark will be publicly released, as this is important for the community to verify and build upon the reported results.
- 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?
The study curated 1482 TAVR patients for a new benchmark. This may need ethical approval for the use of patient data.
- 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 Fig. 1 is overly dense and contains too many texts. Please consider revising it to be more concise and straightforward. It seems that Fig.1 is plotted by image generator, such as NanoBanana, and contains generated typos (“fadid”).
- 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 addressed in this paper is both novel and practical. However, the methodology section requires further clarification to better assess its soundness, and the experimental setup needs more elaboration.
- Reviewer confidence
Very confident (4)
- [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.
Reject
- [Post rebuttal] Please justify your final decision from above.
I thank the authors for their efforts. While the response has partially addressed my concerns, several key issues remain unresolved. Specifically, the response does not adequately address what happens when the risk prediction is incorrect (original review point 2), the process by which the subset of token indices corresponding to clinically salient entities is selected when computing the Token-level Mask (original review point 3), nor does it provide the additional clarification needed for the experimental section (original review point 5). Given these outstanding concerns, I would like to maintain my original score.
Review #3
- Please describe the contribution of the paper
This paper proposes TAVR-VLM, a multimodal framework designed for surgical planning in Transcatheter Aortic Valve Replacement (TAVR). The method integrates 3D CT volumes, echocardiography clips, and tabular clinical data to generate grounded reports and risk predictions. The core technical contribution is the Risk-Conditioned Causal Grounding Attention (R-CGA) module, which implements a top-down hierarchy: predicting global risk, compressing it into a risk bottleneck, and using it to mask visual features to ensure that token-level grounding remains clinically relevant. The approach is evaluated on a custom private M3 TAVR dataset, focusing on reporting quality, risk assessment, and grounding accuracy.
- 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.
Clinical Importance: The application for TAVR planning addresses a highly critical and complex, high-stakes surgical domain, which is less explored compared to general radiology tasks.
Intuitive Design: The “Risk -> Region -> Word” attention hierarchy is clinically well-motivated, aligning the model’s focus with relevant anatomical regions driven by global risk factors.
Heterogeneous Multimodal Integration: The paper makes a solid engineering effort in fusing diverse data types (3D visual, temporal ultrasound clips, and tabular data), which is necessary for comprehensive surgical pathway assessments.
- 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.
Methodological Comparison Fairness: A primary concern is the disparity between the proposed model and its baseline counterparts. TAVR-VLM utilizes a rich set of multimodal inputs (3D CT, Echo, and Tabular data), whereas several compared baselines are generic 2D VLMs or text generators that were not originally designed for such high-dimensional localized input. It is difficult to determine whether the reported performance margin stems from the inherent innovation of the R-CGA module or simply from the significant advantage of having richer input modalities.
Data Transparency: The study relies entirely on a private “M3 TAVR” dataset. The manuscript provides insufficient detail regarding the cohort’s provenance, inclusion criteria, and the specific multi-reader annotation protocols used for establishing ground-truth spatial bounding boxes.
Lack of External Validation: Given the high-risk nature of TAVR surgical planning, the complete lack of external or cross-institutional validation makes the claims regarding the model’s “clinical viability” premature.
Vague Metric Definitions: Certain safety-critical metrics, such as the “Hallucination Rate,” are described vaguely without clear algorithmic definitions of how a generated entity is adjudicated as a hallucination relative to the three input modalities.
- 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 the clinical motivation for multimodal TAVR planning is exceptionally strong and intuitive, the current submission does not meet the necessary standard for methodological rigor and reproducibility. The reliance on an under-described private dataset makes the results difficult to verify. More critically, the comparison between a native high-dimensional multimodal system and generic 2D baselines creates an unfair advantage, making it impossible to assess the true contribution of the R-CGA architecture. Without a level-playing-field baseline comparison and external validation, the claims of clinical readiness are currently unsupported.
- 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.
I appreciate the authors’ detailed response. The rebuttal has addressed my main concerns.
First, regarding baseline fairness, the new ablation study (Table 2) clears this up. Showing that the hallucination rate jumps from 8.2% to 18.4% when removing the R-CGA module (while keeping the exact same multimodal inputs) confirms that the performance gain comes from the proposed architecture, not just from having richer data.
Second, on evaluation objectivity, using expert-annotated ROIs for 200 patients is a solid way to validate the spatial grounding metrics (mIoU). The explanation about the Stop-Gradient operator also resolves my question about early-stage error propagation.
Finally, I am glad the authors agreed to tone down the clinical translation claims and frame the work as an algorithmic proof-of-concept. The promise to open-source the code and provide detailed patient inclusion criteria is also welcome.
Because the technical and evaluation concerns were properly addressed, I am changing my score to Accept. Please make sure all the promised clarifications are actually included in the final version of the manuscript.
Author Feedback
We sincerely thank the Area Chairs and Reviewers for recognizing TAVR-VLM’s clinical importance, intuitive design, and strong performance. We address the primary concerns below.
1.Baseline Fairness and Multimodal Inputs (AC, R3) We respectfully clarify that our experimental design already accounts for this. To maintain domain fairness, we explicitly extracted standardized 2D slices for the 2D-centric baselines. More importantly, the superiority of TAVR-VLM stems from the R-CGA module, not just input dimensionality. We empirically prove this in our ablation study (Table 2): when we use the exact same multimodal inputs but remove the causal visual purification (w/o V_act), the hallucination rate spikes from 8.2% to 18.4%. This isolates the architectural contribution, confirming the performance gap is not an artifact of richer data inputs.
2.Empirical Evidence for Causal Link & Objective Validation (AC, R1, R2, R3) Empirical Evidence: The causal link is empirically substantiated in Table 2 and Figure 2a. When the top-down alignment is severed by disabling the causal consistency loss (lambda_causal=0), spatial grounding precision (mIoU) drops to 0.350 and the hallucination rate doubles to 16.1%. This explicitly proves that enforcing the global risk bottleneck prevents the model from exploiting spurious correlations.
Objective Validation: Our hallucination and spatial anchoring metrics do not rely solely on internal masks. As detailed in Section 3.1, we evaluate spatial grounding precision (mIoU) against independent, expert-annotated region-of-interests (ROIs) obtained for a subset of 200 patients in the test set. This ensures an objective, human-validated assessment of the model’s anatomical anchoring.
3.Methodological Gaps and Training Stability (AC, R2) Definitions: Due to space constraints, some formalisms were condensed. The discrete risk levels (l, m, h) are operationalized based on standard STS scores and VARC-3 guidelines. L_rec is a standard cross-entropy loss applied to the structured recommendation outputs.
Training Stability: R2 correctly notes the risk of early-stage errors propagating. We mitigated this precisely through the Stop-Gradient operator in the causal consistency loss. This asymmetric constraint prevents degenerate solutions where the global mask diffuses to minimize loss. Our ablation confirms that standard backpropagation without Stop-Gradient degrades performance.
4.Dataset Generalizability and Clinical Claims (AC, R1, R3) We agree with R1 that claims of “clinical translation” should be tempered without out-of-distribution external validation; we will revise the manuscript to frame TAVR-VLM as a robust algorithmic proof-of-concept. However, M3TAVR is the largest structurally annotated cohort of its kind (1,482 patients). To ensure internal validity and prevent data leakage, we enforced a strict patient-level stratified split (70/10/20). Because comprehensive 3D+Temporal+Tabular TAVR datasets do not publicly exist, we rely on this rigorous internal validation. We commit to clarifying our cohort inclusion criteria and open-sourcing our core codebase to facilitate community reproducibility.
We sincerely appreciate the committee’s rigorous feedback. By clarifying our domain-fair baselines, emphasizing the objective ROI validation, and empirically substantiating the causal link, we hope to have resolved the primary concerns. We believe TAVR-VLM represents a significant, fundamentally interpretable step forward for surgical AI, and we hope these clarifications merit acceptance at MICCAI.
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.
While reviewers agreed that the paper addresses a clinically critical and under-explored task with a well-motivated, interpretable framework aligned with cardiologists’ workflow, they also raised serious concerns about the paper’s evaluation rigor, methodological clarity, and generalizability: First, experiments are limited to a single private dataset without external or cross-institutional validation, and metric definitions lack objectivity and clarity. Second, methodological comparisons are unfair, as baselines lack the rich multimodal inputs used by the proposed model. Third, key methodological aspects are insufficiently described or justified, including undefined risk levels, unclear training stability for the risk estimator, and unsubstantiated claims about the causal link underlying hallucinations. Since the reviews are mixed, we would like to invite the authors to provide a rebuttal. In addition to clarify other concerns, please try to address the following points during rebuttal: if possible, validate the model’s generalizability with external/multi-center data or public datasets; ensure fair baseline comparisons by adapting baselines to use the same multimodal inputs; clarify methodological gaps and justify key choices; and provide empirical/theoretical evidence for the causal link between risk states and hallucinations, along with objective validation of the hallucination rate.
- 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 is a borderline paper with mixed reviews. Reviewers agreed that the work addresses a clinically critical and underexplored task in TAVR preoperative planning and proposes a structurally interpretable R-CGA framework that aligns well with cardiologists’ workflows. However, they also raised serious concerns about the paper’s soundness and significance: first, the reported metrics such as CIDEr may be inflated due to the template-based nature of TAVR reports, and the definition of hallucination relies heavily on internal model signals rather than independent clinical validation; second, the experimental validation is limited to a single private dataset, with insufficient transparency regarding cohort selection, annotation protocols, and no external or multi-center validation; third, key methodological details remain unclear, including the handling of incorrect risk predictions, the selection of clinically salient tokens, and the fairness of comparisons given the model’s access to richer multimodal inputs than baselines. The rebuttal addressed some concerns, but did not resolved all. After careful deliberation, I lean a bit towards accepting this manuscripts due to its potential clinical values.
Meta-review #2
- After you have reviewed the rebuttal and updated reviews, please provide your recommendation based on all reviews and the authors’ rebuttal.
Reject
- Please justify your recommendation.
Despite authors’ efforts to address the concerns of all three reviewers, some still remained unaddressed. As Reviewer 2 specifies, when the risk prediction is incorrect, it is still unclear how the system responds. On the other hand, given lack of external validation, the hallucination evaluation lacks objectivity.
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.
Although one reviewer maintained concerns regarding incomplete clarification of risk-prediction failure cases, token-level mask construction, and experimental details, the other reviewers provided very positive assessments and found the rebuttal substantially strengthened the paper. In particular, the authors added convincing ablation evidence isolating the contribution of the proposed module, clarified the use of human expert ground truth for spatial evaluation, agreed to moderate clinical claims, and committed to code release. Given the strong post-rebuttal support from two reviewers and the paper’s value as a useful proof-of-concept, I recommend acceptance, while encouraging the authors to address the remaining clarity issues in the final version.
