TAVR-VLM: Risk-Conditioned Causal Grounding for Hallucination-Resistant Report Generation

2026-06-25

AI Summary

This work addresses the issue of diagnostic hallucinations in multimodal large language models used for transcatheter aortic valve replacement (TAVR) planning, which arise from insufficient anatomical grounding. To mitigate this, the authors propose a Risk-Conditioned Causal Grounding Attention (R-CGA) mechanism that establishes a causal pathway from “risk → region → word,” compressing multimodal inputs into a global risk mask during autoregressive generation. A projection consistency objective enforces alignment between textual descriptions and imaging-derived risk regions. Evaluated on a cohort of 1,482 patients, the model achieves an AUROC of 0.896 and a CIDEr score of 0.936, while reducing hallucination rates to 8.1%. These results demonstrate significantly enhanced anatomical interpretability and clinical credibility of generated reports, establishing a new state-of-the-art performance for this task.

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 $\text{M}^3\text{TAVR}$, a comprehensive 1,482-patient cohort, TAVR-VLM establishes a new state-of-the-art. It achieves an AUROC of 0.896, boosts CIDEr to 0.936, and drastically reduces the hallucination rate to 8.1\%, thereby improving interpretability for evidence-based surgical AI.

Problem

• hallucination
• multimodal reasoning
• anatomical grounding
• TAVR
• diagnostic reliability

Innovation

• Risk-Conditioned Causal Grounding
• Multimodal Large Language Models
• Hallucination Reduction
• Structural Grounding Pathway
• Transcatheter Aortic Valve Replacement