MedDr_0401

MedDr: Diagnosis-Guided Bootstrapping for Large-Scale Medical Vision-Language Learning

A generalist foundation model for healthcare capable of handling diverse medical data modalities.

<img src="https://github.com/sunanhe/MedDr/raw/main/examples/logo.jpg" width="200"/>

Authors: Sunan He*, Yuxiang Nie*, Zhixuan Chen, Zhiyuan Cai, Hongmei Wang, Shu Yang, Hao Chen**
(*Equal Contribution, **Corresponding Author)
Institution: SMART Lab, Hong Kong University of Science and Technology

---

Model Summary

MedDr is a large-scale generalist vision-language model for healthcare. It is built upon InternVL and trained using a diagnosis-guided bootstrapping strategy that leverages both image and label information to construct high-quality vision-language datasets.

MedDr supports diverse medical imaging modalities:

• 🫁 Radiology (X-ray, CT, MRI)
• 🔬 Pathology
• 🧴 Dermatology
• 👁️ Retinography
• 🔭 Endoscopy

During inference, MedDr employs a retrieval-augmented medical diagnosis strategy to enhance generalization ability.

---

Capabilities

• Visual Question Answering (VQA) for medical images
• Medical report generation
• Medical image diagnosis across multiple modalities

---

Usage

Environment Setup

This model is built on InternVL. Please follow the INSTALLATION.md to set up the environment.

Quick Demo

# Clone the GitHub repository
# git clone https://github.com/sunanhe/MedDr.git

# Edit demo.py and set model_path to your local checkpoint directory
# Then run:
# python3 demo.py

See demo.py in the GitHub repository for a full example.

---

Citation

If you find MedDr useful in your research, please consider citing:

@article{he2024meddr,
  title={MedDr: Diagnosis-Guided Bootstrapping for Large-Scale Medical Vision-Language Learning},
  author={He, Sunan and Nie, Yuxiang and Chen, Zhixuan and Cai, Zhiyuan and Wang, Hongmei and Yang, Shu and Chen, Hao},
  journal={arXiv preprint arXiv:2404.15127},
  year={2024}
}

---

Acknowledgements

This work builds upon InternVL. We thank the InternVL team for their outstanding contributions to the open-source VLM community.