SSL4DSA

This repository is the Python (Pytorch) implement of paper 'Yao Pu, et al., Semi-supervised Segmentation of Coronary DSA using Mixed Networks and Multi-strategies'(https://www.sciencedirect.com/science/article/pii/S001048252201201X). Semi-Supervised Learning (SSL), Inception Swin-Unet (ISUnet), Pyramid consistency Learning (PL), & Confidence Learning (CL) are adopted in to segment the coronary artery of human in Digital Subtraction Angiography (DSA) images.

@article{pu2022semi,
  title={Semi-supervised segmentation of coronary DSA using mixed networks and multi-strategies},
  author={Pu, Yao and Zhang, Qinghua and Qian, Cheng and Zeng, Quan and Li, Na and Zhang, Lijuan and Zhou, Shoujun},
  journal={Computers in Biology and Medicine},
  pages={106493},
  year={2022},
  publisher={Elsevier}
}

0. Framework

1. Requirements

Python: 3.8.10
torch: 1.11.0+cu113
cleanlab: 2.0.0
tensorboardX: 2.5
tqdm: 4.64.0
einops: 0.4.1
timm: 0.5.4
ml-collections: 0.1.1
...

2. Contents

.
│  .gitignore
│  README.md
│
├─dada
│  └─CA_DSA
│      ├─LCA
│      │  │  test.list
│      │  │  train.list
│      │  │
│      │  ├─data
│      │  │     000.bmp
│      │  │     ...
│      │  │     150.bmp
│      │  │  
│      │  ├─enhance
│      │  ├─label
│      │  │     000.bmp
│      │  │     ...
│      │  │     150.bmp
│      │  └─skeleton
│      └─RCA
│          │  test.list
│          │  train.list
│          │
│          ├─data
│          │    000.bmp
│          │    ...
│          │    150.bmp
│          ├─enhance
│          ├─label
│          │    000.bmp
│          │    ...
│          │    150.bmp
│          └─skeleton
│  
├─code
│  │  dataset.py                                    # load dataset
│  │  histogram.ipynb
│  │  maxpool.ipynb
│  │  preprocess.py                                 # CLAHE preprocessing
│  │  train_supervised_softmax.py                   # FSL OneNetwork
│  │  train_semisupervised_UAMT.py                  # SSL TwoSameNetworks UAMT
│  │  train_semisupervised_CNN_Transformer.py       # SSL CNN_SwinUnet
│  │  train_semisupervised_CNN_Transformer_CL.py    # SSL CNN_ISUnet_CL
│  │  train_semisupervised_CNN_Transformer_PL.py    # SSL CNN_ISUnet_PL
│  │  train_semisupervised_CNN_Transformer_PLCL.py  # SSL CNN_ISUnet_PLCL: the MAIN training file
│  │  test_model.py                                 # the testing file for single_out_network
│  │  test_model_multiout.py                        # the MAIN testing file for multi_out_network
│  │  run.sh
│  │  
│  ├─networks
│  │  │  __init__.py
│  │  │  Unet.py
│  │  │  Unet_4out.py
│  │  │  Vnet.py
│  │  │  Vnet_4out.py
│  │  │  Unetpp.py
│  │  │  AttUnet.py
│  │  │
│  │  ├─SwinUnet
│  │  │      swin_transformer_unet_skip_expand_decoder_sys.py
│  │  │      vision_transformer.py
│  │  │
│  │  └─InceptionSwinUnet
│  │          InceptionSwinUnet_v1.py
│  │          InceptionSwinUnet_v2.py
│  │          InceptionSwinUnet_v2_4out.py
│  │          InceptionSwinUnet_v3.py
│  │          InceptionSwinUnet_v3_4out.py           # the MAIN network (ISUnet) file
│  │
│  └─utils
│          generateCenterline.py
│          losses.py
│          ramps.py
│  
├─model
│    ├─predictions
│    ├─semisupervised
│    └─supervised
│
└─img
       CLAHE.png
       Framework.png
       ISUnet.png
       ITM.png
       Result.png

3. Steps

3.1. Pre-Processing

code/prepocess.py enhance raw DSA by approach CHALE (Contrast Limited Adaptive Histogram Equalization).

3.2. Dataloader

code/dataset.py load the data & do some transform.

3.3. Networks

code/networks/ includes all the networks we used in the experiments.

• Unet
• Unet_4out
• Vnet
• Vnet_4out
• Unet++
• AttentionUnet
• SwinUnet
• ISUnetV1
• ISUnetV2
• ISUnetV3
• ISUnetV2_4out
• ISUnetV3_4out

3.4. Training

De-comment training lines in the code/run.sh and run it.

cd code
bash run.sh

Or run the python file directly.

cd code

# left coronary artery (LCA)
python train_semisupervised_CNN_Transformer_PLCL.py --root_path='../data/CA_DSA/LCA/' --exp='semisupervised/CNN_Transformer_PLCL/unet4out_isunetv5/LCA' --nets='unet4out_isunetv5' --batch_size=2 --gpus='1,2' --labeled_num=20 

# right coronary artery (RCA)
python train_semisupervised_CNN_Transformer_PLCL.py --root_path='../data/CA_DSA/RCA/' --exp='semisupervised/CNN_Transformer_PLCL/unet4out_isunetv5/RCA' --nets='unet4out_isunetv5' --batch_size=2 --gpus='1,2' --labeled_num=20 

3.5. Testing

De-comment Testing lines in the code/run.sh and run it.

cd code
bash run.sh

Or run the python file directly.

cd code

# load LCA model1
python test_model_multiout.py --root_path='../data/CA_DSA/LCA/' --model='semisupervised/CNN_Transformer_PLCL/unet4out_isunetv5/LCA' --whichmodel='model1'

# load LCA model2
python test_model_multiout.py --root_path='../data/CA_DSA/LCA/' --model='semisupervised/CNN_Transformer_PLCL/unet4out_isunetv5/LCA' --whichmodel='model2'

# load RCA model1
python test_model_multiout.py --root_path='../data/CA_DSA/RCA/' --model='semisupervised/CNN_Transformer_PLCL/unet4out_isunetv5/RCA' --whichmodel='model1'

# load RCA model2
python test_model_multiout.py --root_path='../data/CA_DSA/RCA/' --model='semisupervised/CNN_Transformer_PLCL/unet4out_isunetv5/RCA' --whichmodel='model2'

4.Result