Dual-task Consistency

Code for this paper: Semi-supervised Medical Image Segmentation through Dual-task Consistency (AAAI2021)

@inproceedings{luo2021semi,
  title={Semi-supervised Medical Image Segmentation through Dual-task Consistency},
  author={Luo, Xiangde and Chen, Jieneng and Song, Tao and Wang, Guotai},
  booktitle={Proceedings of the AAAI Conference on Artificial Intelligence},
  volume={35},
  number={10},
  pages={8801--8809},
  year={2021}
}

Requirements

Some important required packages include:

• Pytorch version >=0.4.1.
• TensorBoardX
• Python == 3.6
• Some basic python packages such as Numpy, Scikit-image, SimpleITK, Scipy ......

Follow official guidance to install Pytorch.

Usage

1. Clone the repo:

git clone https://github.com/HiLab-git/DTC.git 
cd DTC

2. Put the data in data/2018LA_Seg_Training Set.

3. Train the model

cd code
python train_la_dtc.py

4. Test the model

python test_LA.py

Our pre-trained models are saved in the model dir DTC_model (both 8 labeled images and 16 labeled images), and the pretrained SASSNet and UAMT model can be download from SASSNet_model and UA-MT_model. The other comparison method can be found in SSL4MIS

Results on the Left Atrium dataset (SOTA).

• The training set consists of 16 labeled scans and 64 unlabeled scans and the testing set includes 20 scans.

Methods	DICE (%)	Jaccard (%)	ASD (voxel)	95HD (voxel)	Reference	Released Date
UAMT	88.88	80.21	2.26	7.32	MICCAI2019	2019-10
SASSNet	89.54	81.24	2.20	8.24	MICCAI2020	2020-07
DTC	89.42	80.98	2.10	7.32	AAAI2021	2020-09
LG-ER-MT	89.62	81.31	2.06	7.16	MICCAI2020	2020-10
DUWM	89.65	81.35	2.03	7.04	MICCAI2020	2020-10
MC-Net	90.34	82.48	1.77	6.00	Arxiv	2021-03

• The training set consists of 8 labeled scans and 72 unlabeled scans and the testing set includes 20 scans.

Methods	DICE (%)	Jaccard (%)	ASD (voxel)	95HD (voxel)	Reference	Released Date
UAMT	84.25	73.48	3.36	13.84	MICCAI2019	2019-10
SASSNet	87.32	77.72	2.55	9.62	MICCAI2020	2020-07
DTC*	87.51	78.17	2.36	8.23	AAAI2021	2020-09
LG-ER-MT	85.54	75.12	3.77	13.29	MICCAI2020	2020-10
DUWM	85.91	75.75	3.31	12.67	MICCAI2020	2020-10
MC-Net	87.71	78.31	2.18	9.36	Arxiv	2021-03

• Note that, * denotes the results from MC-Net and the model has been openly available (provided by Dr. YiCheng), thanks for Dr. Yicheng.

Acknowledgement

• This code is adapted from UA-MT, SASSNet, SegWithDistMap.
• We thank Dr. Lequan Yu, M.S. Shuailin Li and Dr. Jun Ma for their elegant and efficient code base.
• More semi-supervised learning approaches for medical image segmentation have been summarized in SSL4MIS.