SAMUS

This repo is the official implementation for:\ SAMUS: Adapting Segment Anything Model for Clinically-Friendly and Generalizable Ultrasound Image Segmentation.\ (The details of our SAMUS can be found in the models directory in this repo or in the paper.)

Highlights

🏆 Low GPU requirements. (one 3090ti with 24G GPU memory is enough)\ 🏆 Large ultrasound dataset. (about 30K images and 69K masks covering 6 categories)\ 🏆 Excellent performance, especially in generalization ability.

Installation

Following Segment Anything, python=3.8.16, pytorch=1.8.0, and torchvision=0.9.0 are used in SAMUS.

1. Clone the repository.

   git clone https://github.com/xianlin7/SAMUS.git
   cd SAMUS

2. Create a virtual environment for SAMUS and activate the environment.

   conda create -n SAMUS python=3.8
   conda activate SAMUS

3. Install Pytorch and TorchVision. (you can follow the instructions here)
4. Install other dependencies.

   pip install -r requirements.txt

   Checkpoints

   We use checkpoint of SAM in vit_b version.

Data

• US30K consists of seven publicly-available datasets, including TN3K, DDTI, TG3K, BUSI, UDIAT, CAMUS, and HMC-QU.
• All images were saved in PNG format. No special pre-processed methods are used in data preparation.
• We have provided some examples to help you organize your data. Please refer to the file fold example_of_required_dataset_format.\ Specifically, each line in train/val.txt should be formatted as follows:

  <class ID>/<dataset file folder name>/<image file name>

• The relevant information of your data should be set in ./utils/config.py

Training

Once you have the data ready, you can start training the model.

cd "/home/...  .../SAMUS/"
python train.py --modelname SAMUS --task <your dataset config name>

Testing

Do not forget to set the load_path in ./utils/config.py before testing.

python test.py --modelname SAMUS --task <your dataset config name>

Citation

If our SAMUS is helpful to you, please consider citing:

@misc{lin2023samus,
      title={SAMUS: Adapting Segment Anything Model for Clinically-Friendly and Generalizable Ultrasound Image Segmentation}, 
      author={Xian Lin and Yangyang Xiang and Li Zhang and Xin Yang and Zengqiang Yan and Li Yu},
      year={2023},
      eprint={2309.06824},
      archivePrefix={arXiv},
      primaryClass={cs.CV}
}