This repo is the official implementation of CS-Net: Channel and Spatial Attention Network for Curvilinear Structure Segmentation.
The main contribution of this work is the publication of two scarce datasets in the medical image field. Plesae click the link below to access the details and source data.
Using the train.py
and predict.py
to train and test the model on your own dataset, respectively.
- PyTorch >= 0.4.1
- tqdm
- cv2
- visdom
- sklearn
The attention module was implemented based on DANet. The difference between the proposed module and the original block is that we added a new 1x3 and 3x1 kernel convolution layer into spatial attention module. Plese refer to the paper for details.
@inproceedings{mou2019cs,
title={CS-Net: channel and spatial attention network for curvilinear structure segmentation},
author={Mou, Lei and Zhao, Yitian and Chen, Li and Cheng, Jun and Gu, Zaiwang and Hao, Huaying and Qi, Hong and Zheng, Yalin and Frangi, Alejandro and Liu, Jiang},
booktitle={International Conference on Medical Image Computing and Computer-Assisted Intervention},
pages={721--730},
year={2019},
organization={Springer}
}
The extension of the 2D CS-Net
train3d.py
is used to train the 3D segmentation network.
predict3d.py
is used to test the trained model.
Please note that you should change the dataloader definition in train3d.py
.
Requirements:
-
PyTorch = 0.4.1
-
visdom
-
SimpleITK:
pip install SimpleITK
@article{mou2020cs2, title={CS2-Net: Deep Learning Segmentation of Curvilinear Structures in Medical Imaging}, author={Mou, Lei and Zhao, Yitian and Fu, Huazhu and Liux, Yonghuai and Cheng, Jun and Zheng, Yalin and Su, Pan and Yang, Jianlong and Chen, Li and Frangi, Alejandro F and others}, journal={Medical Image Analysis}, pages={101874}, year={2020}, publisher={Elsevier} }
II-4 Correction to: CS2-Net- Deep learning segmentation of curvilinear structures in medical imaging
The original comparison results in Table 8 on page 14 are:
The corrected comparison results are: