TAVR-VLM Reduces Hallucinations in Medical Report Generation

Multimodal Large Language Models (MLLMs) hold great promise for complex medical tasks like Transcatheter Aortic Valve Replacement (TAVR) planning, which requires integrating diverse data. However, their deployment is hindered by diagnostic hallucinations, where generated text lacks accurate anatomical grounding. To overcome this, researchers introduce TAVR-VLM, a new framework specifically engineered to resist such hallucinations. TAVR-VLM incorporates a mechanism called Risk-Conditioned Causal Grounding Attention (R-CGA). This creates an internal "Risk → Region → Word" structural grounding pathway, effectively compressing multimodal inputs into a causal risk bottleneck. This process purifies visual features into a global risk mask. During text generation, a support-projected causal consistency objective ensures that token-level grounding remains within the risk-defined support mask. Evaluated on a large patient cohort, TAVR-VLM achieved state-of-the-art results, significantly boosting performance metrics and reducing the hallucination rate to a mere 8.1%, thereby enhancing the interpretability of AI in surgical planning.