TransUNet

This repo holds code for TransUNet: Transformers Make Strong Encoders for Medical Image Segmentation

📰 News

• [10/15/2023] 🔥 3D version of TransUNet is out! Our 3D TransUNet surpasses nn-UNet with 88.11% Dice score on the BTCV dataset and outperforms the top-1 solution in the BraTs 2021 challenge. Please take a look at the code and paper.

Usage

1. Download Google pre-trained ViT models

• Get models in this link: R50-ViT-B_16, ViT-B_16, ViT-L_16...

wget https://storage.googleapis.com/vit_models/imagenet21k/{MODEL_NAME}.npz &&
mkdir ../model/vit_checkpoint/imagenet21k &&
mv {MODEL_NAME}.npz ../model/vit_checkpoint/imagenet21k/{MODEL_NAME}.npz

2. Prepare data

Please go to "./datasets/README.md" for details, or use the preprocessed data and data2 for research purposes.

3. Environment

Please prepare an environment with python=3.7, and then use the command "pip install -r requirements.txt" for the dependencies.

4. Train/Test

• Run the train script on synapse dataset. The batch size can be reduced to 12 or 6 to save memory (please also decrease the base_lr linearly), and both can reach similar performance.

CUDA_VISIBLE_DEVICES=0 python train.py --dataset Synapse --vit_name R50-ViT-B_16

• Run the test script on synapse dataset. It supports testing for both 2D images and 3D volumes.

python test.py --dataset Synapse --vit_name R50-ViT-B_16

Reference

• Google ViT
• ViT-pytorch
• segmentation_models.pytorch

Citations

@article{chen2021transunet,
  title={TransUNet: Transformers Make Strong Encoders for Medical Image Segmentation},
  author={Chen, Jieneng and Lu, Yongyi and Yu, Qihang and Luo, Xiangde and Adeli, Ehsan and Wang, Yan and Lu, Le and Yuille, Alan L., and Zhou, Yuyin},
  journal={arXiv preprint arXiv:2102.04306},
  year={2021}
}