VT-UNet/version_1

• VT-UNet
• • Environment
  • Data Preparation
  • Pre-Trained Weights
  • Pre-Trained Base Model For BraTS 2021
  • Train/Test
  • Acknowledgements
  • References
  • Citing VT-UNet

VT-UNet

This repo contains the supported pytorch code and configuration files to reproduce 3D medical image segmentaion results of VT-UNet.

Environment

Prepare an environment with python=3.8, and then run the command "pip install -r requirements.txt" for the dependencies.

Data Preparation

• For experiments we used four datasets:

  • BRATS 2021 : http://braintumorsegmentation.org/
  • MSD BRATS, LIVER, PANCREAS : http://medicaldecathlon.com/

• File structure

   BRATS2021
    |---Data
    |   |--- RSNA_ASNR_MICCAI_BraTS2021_TrainingData
    |   |   |--- BraTS2021_00000
    |   |   |   |--- BraTS2021_00000_flair...
    |   
    |              
    |   
    |
   VT-UNet
    |---train.py
    |---test.py
    |---pretrained_ckpt
    |---saved_model
    ...
  

Pre-Trained Weights

• Swin-T: https://github.com/SwinTransformer/storage/releases/download/v1.0.0/swin_tiny_patch4_window7_224.pth
• Download Swin-T pre-trained weights and add it under pretrained_ckpt folder

Pre-Trained Base Model For BraTS 2021

• VT-UNet-B: https://drive.google.com/file/d/1tLpEfyKgQ8xgvM3D1Aqi8aD3ILVB3du7/view?usp=sharing
• Download VT-UNet-B pre-trained model and add it under saved_model folder before running test.py

Train/Test

• Train : Run the train script on BraTS 2021 Training Dataset with Base model Configurations.

python train.py --cfg configs/vt_unet_base.yaml --num_classes 3 --epochs 350

• Test : Run the test script on BraTS 2021 Training Dataset.

python test.py --cfg configs/vt_unet_base.yaml --num_classes 3

Acknowledgements

This repository makes liberal use of code from open_brats2020, Swin Transformer, Video Swin Transformer and Swin-Unet

References

• BraTS 2021
• Medical segmentation decathlon

Citing VT-UNet

    @misc{peiris2021volumetric,
      title={A Volumetric Transformer for Accurate 3D Tumor Segmentation}, 
      author={Himashi Peiris and Munawar Hayat and Zhaolin Chen and Gary Egan and Mehrtash Harandi},
      year={2021},
      eprint={2111.13300},
      archivePrefix={arXiv},
      primaryClass={eess.IV}
    }