Skip to content
/ DTC Public

Semi-supervised Medical Image Segmentation through Dual-task Consistency

License

Notifications You must be signed in to change notification settings

HiLab-git/DTC

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

96 Commits
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

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

  2. Train the model

cd code
python train_la_dtc.py
  1. 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.

About

Semi-supervised Medical Image Segmentation through Dual-task Consistency

Topics

Resources

License

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published

Languages