GCtx-UNet: Efficient Deep Network for Medical Image Segmentation

GCtx-UNet is a U-shaped network architecture that incorporates the Global Context Vision Transformer (GC-ViT) to enhance medical image segmentation by effectively capturing both global and local features. Feel free to check out our preprint on arXiv. ⁣

Pre-trained model:

1. Download pre-trained GC-ViT transformer model (GCViT-xxtiny) pre-trained on ImageNet1K:

[Get pre-trained model in this link] (https://drive.usercontent.google.com/download?id=1apSIWQCa5VhWLJws8ugMTuyKzyayw4Eh&export=download&authuser=0): Put pretrained xx-Tiny into folder "pretrained_ckpt/"

2. Download pre-trained GC-ViT transformer model (GCViT-xxtiny) pre-trained on MedNet:

• Get pre-trained GCViT-xxTiny on MedNet :Put pretrained xx-Tiny into folder "pretrained_ckpt/"

Prepare data

The Synapse and ACDC datasets we used are provided by the authors of TransUnet. You can access the processed data through this link. For more details, please refer to the "./datasets/README.md" file. Alternatively, you can go ahead and request the preprocessed data by emailing [email protected]. If you use the preprocessed data, please ensure it is solely for research purposes and do not redistribute it by TransUnet's License.

Environment

• Please prepare an environment with python=3.8, and then use the command "pip install -r requirements.txt" for the dependencies.

Train/Test

1) Train

CUDA_VISIBLE_DEVICES=0  python -W ignore train.py 

2) Testing

Get pre-trained GCtx-UNet model weights for the Synapse dataset: link

download the file and put it into the folder model_out.

CUDA_VISIBLE_DEVICES=0  python -W ignore test.py --is_saveni

References

• Swin-Unet
• TransUnet
• GC-ViT

Citation

@article{alrfou2024gctx,
  title={GCtx-UNet: Efficient Network for Medical Image Segmentation},
  author={Alrfou, Khaled and Zhao, Tian},
  journal={arXiv preprint arXiv:2406.05891},
  year={2024}
}