The official code for the paper "Towards Generalist Foundation Model for Radiology by Leveraging Web-scale 2D&3D Medical Data"

ArXiv

Website

Model checkpoint

In this project, we collect a large-scale medical multi-modal dataset, MedMD, with 16M 2D or 3D images. We train a new medical multi-modal generative model RadFM on it, enabling both 2D and 3D scans, multi-image input and visual-language interleaving cases.

For quick start, you can check the Quick_demo path.
We demonstrate a simple diagnosis case here to show how to inference with our model.
Feel free to modify it as you want.

• S1. Download Model checkpoint or form baiduyun (No need for decompressing).

• S2. Decompress the original zip file, you can get a pytorch_model.bin.

• S3. put pytorch_model.bin under path Quick_demo/.

• S4. python test.py and you can get a conversation as:

  Input: Can you identify any visible signs of Cardiomegaly in the image?
  Output: yes

By the way, never try to perform this in cpu and gpu is all you need :）.

For re-training a model on our dataset or large-scale testing our pre-train model, you can check src.

Simply, train.py for training and test.py for testing.

• Check the data_csv to get how different datasets are processed and download them into src/Dataset/data_csv
• Modify the path as you disire, and check src/train.py to pre-train or src/train.py to test.

Some cases produced by our final model:

MedKD Dataset downloading URL:

• Polish the code in src for eaiser reading.
• upload huggingface version.

We sincerely thank all the contributors who uploaded the relevant data in our dataset online. We appreciate their willingness to make these valuable cases publicly available.

If you have any questions, please feel free to contact [email protected].