CCD

关于

Multimodal large language models (MLLMs) are advancing radiology by combining image and text understanding, but often generate inaccurate or unsupported clinical details—so-called medical hallucinations. Clinical Contrastive Decoding (CCD) is a training-free and retrieval-free inference framework that integrates structured clinical signals from task‑specific radiology expert models. CCD reduces these hallucinations and improves clinical accuracy without changing the base MLLM, offering a practical way to make radiology MLLMs more reliable.

主要功能

• 5 GitHub stars
• Compatible with Libra/LLaVA-architecture radiology MLLMs and multiple expert models
• Enhances clinical accuracy without modifying the base MLLM architecture
• Training-free and retrieval-free inference framework
• Integrates structured clinical signals from radiology expert models
• Mitigates medical hallucinations in MLLM-generated radiology reports

使用案例

• Provides an educational tool for demonstrating medical image interpretations and temporal progressions
• Supports radiologists in drafting and reviewing chest X-ray reports for clinical decision making
• Serves as a research utility for investigating automated report generation and visual-language alignment in medical imaging