This is an official pytorch implementation of the paper

Our framework consists of two steps, 1) Image translation and 2) Self-training. In this repository, only step 1 is supported. *And we are currently building on it!

(worked on the setting below, and do not guarantee other versions)

argparse                1.4.0
numpy                   1.22.3
pytorch                 1.11.0
torchvision             0.12.0
tensorboard             2.9.1
einops                  0.4.1
scikit-learn            1.1.1
scipy                   1.8.1
cudatoolkit             11.3.1
python                  3.9.12
simpleitk               2.0.2

(and need to install extra dependencies required. Information about the packages will be updated!)

You need to modify the data and fit to datalist format provided. We splitted 3D medcial data to every single slice with file, for efficient loading.

The dataset used for training can be downloaded here for each.

CrossMoDA (Vestibular and schwannoma) : https://crossmoda-challenge.ml/

Cardiac (MMWHS) : https://zmiclab.github.io/zxh/0/mmwhs/

python train.py

You can check the status using

tensorboard --logdir tensorboard/<Your_Experiment_Name>

and your checkpoint is saved at /checkpoints.

This code was partially borrowed by

https://github.com/Seung-Hun-Lee/DRANet

https://github.com/lucidrains/segformer-pytorch

and about the nnU-Net part with Sensitivity & Specificity aware pseudo-label refinement algorithm is to be updated!