The paper “Hallucination-Aware Multimodal Benchmark for Gastrointestinal Image Analysis with Large Vision-Language Models” has been accepted at MICCAI 2025!
It tackles the critical issue of hallucination in medical vision-language models (VLMs), where the generated descriptions are inconsistent with the visual content, posing serious risks in clinical settings. To address this, the authors introduce Gut-VLM, a novel multimodal dataset focused on gastrointestinal imaging. In addition, the paper also establishes a new benchmark by evaluating state-of-the-art VLMs across multiple metrics, offering a valuable resource for advancing safe and accurate medical AI.
