This is a stand-alone self-contained Python project to train, run, and evaluate the SwinUNETR model.

Prerequisite

• Install conda.

Prepare

First prepare a Python virtual environment by

conda create --name SwinUNETR python=3.11
conda activate SwinUNETR
pip install -r requirements.txt --extra-index-url https://download.pytorch.org/whl/cu117

Test your environment by running

python -c "import torch; print(torch.cuda.device_count())"

which should show the correct number of GPUs.

Then create a local env file by

cp .env .env.local

and set SWINUNETR_WORKSPACE to be the path of some directory with sufficiently large space, and SWINUNETR_DATA_ROOT to be the uncompressed folder of the BraTS2021 dataset containing subfolders like

- BraTS2021_00001/
- BraTS2021_00002/
- BraTS2021_00003/
...

Pipeline

The codebase provides a reasonable default setting. Run following commands in turn to train, predict, and evaluate.

make train
make predict
make evaluate