Releases: X-iZhang/CCD
Releases · X-iZhang/CCD
v1.0.0
📦 CCD v1.0.0 Release Notes
🎉 Initial Release
We're excited to announce the first official release of CCD (Clinical Contrastive Decoding) — a training-free, retrieval-free inference framework that mitigates hallucinations in radiology multimodal large language models (MLLMs).
🚀 What's Included
- ✅ Clinical Contrastive Decoding Algorithm — reduces medical hallucinations by integrating structured clinical signals from expert models through dual-branch contrastive decoding.
- 🩺 Expert Model Integration — supports DenseNet (CheXpert) and MedSiglip for clinical label extraction from chest X-rays.
- 🤖 7 Pre-configured MLLM Models — including Libra-v1.0-7B, Libra-v1.0-3B, MAIRA-2, LLaVA-Med-v1.5, LLaVA-Rad, Med-CXRGen-F, and Med-CXRGen-I.
- 🎮 Interactive Gradio Demo — available locally and on Hugging Face Spaces.
- 🗂️ Preprocessed Evaluation Datasets — MIMIC-CXR, IU-Xray, CheXpert Plus, and Medical-CXR-VQA test splits on Hugging Face Collections.
- 🛠️ Flexible Python API —
ccd_eval()for CCD inference andrun_eval()for baseline comparison. - 👁️ View Classification Model — identifies and standardises view types (e.g., Frontal, Lateral).
📖 Resources
- 🔍 See the Installation Guide for setup instructions.
- ⚡ Check the Quick Start Guide for CLI, script, and web interface usage.
- 🛠️ Explore Advanced Usage for model switching and parameter tuning.
- 📊 Use RadEval for comprehensive evaluation metrics.
- 📄 Read our arXiv preprint for technical details.
⚠️ Important Note
CCD is intended for research and educational purposes only. All outputs must be reviewed by qualified medical professionals before informing any clinical decision.