Anatomical Region-Guided Contrastive Decoding: A Plug-and-Play Strategy for Mitigating Hallucinations in Medical VLMs

This article introduces a novel method to improve the reliability of medical Visual Language Models (VLMs) by addressing the issue of hallucinations. The approach, "Anatomical Region-Guided Contrastive Decoding," is presented as a plug-and-play strategy, suggesting ease of implementation. The focus on medical applications highlights the importance of accuracy in this domain. The use of contrastive decoding is a key aspect, likely involving comparing different outputs to identify and mitigate errors. The source being ArXiv indicates this is a pre-print, suggesting the work is under review or recently completed.