CCM-stainGAN

---

This code is for the virtual staining part of paper Resection-Inspired Histopathological Diagnosis of Cerebral Cavernous Malformations Using Quantitative Multiphoton Microscopy.

• 2023.09.15 🔥 The related work of CCM-StainGAN has been published by Nature Communications (paper).
• 2022.11.02 🎉 The article was selected as the cover of the journal.
• 2022.08.30 🚀 Update the code of the repo.

Preview

2 Channels	2 Channels	3 Channels

1. Create Envirement:

• Python 3.6 (Recommend to use Anaconda)

• NVIDIA GPU + CUDA

• PyTorch >= 1.8.0

• Python packages:

  cd CCM-stainGAN
  
  conda activate CCM-stainGAN
  
  pip install -r requirements.txt

2. Data Preparation

• The MPM and H&E images are available, download from (Google Drive / Baidu Disk). Other data used in this repo are available from the corresponding author upon reasonable request.

  Note: access code for Baidu Disk is ccm1.

• Split the images into patches.

• Place the training MPM patches to /CCM-stainGAN/dataset/trian/A/ and the H&E patches to /CCM-stainGAN/dataset/trian/B/.

• Place the training MPM category patches to /CCM-stainGAN/dataset/trian/cls_A and the H&E category patches to /CCM-stainGAN/dataset/trian/cls_B .

• Place the testing MPM patches to /CCM-stainGAN/dataset/test/A/ and the H&E patches to /CCM-stainGAN/dataset/test/B/.

• Then this repo is collected as the following form:

  |--CCM-stainGAN
    |--train.py
    |--utils.py
    |--data_utils.py
    |--test.py
    |--output
    |--dataset
      |--MPM2HE
        |--train
          |--A
            |--mpm_000.png
            |--mpm_001.png
            :
            |--mpm_099.png
          |--B
            |--he_000.png
            |--he_001.png
            :
            |--he_099.png
          |--cls_A
            |--he_001_0.png
            |--he_002_0.png
            :
            |--he_099_8.png
          |--cls_B
            |--mpm_001_0.png
            |--mpm_002_0.png
            :
            |--mpm_099_8.png
        |--test
          |--A
            |--mpm_000.png
            :
            |--mpm_099.png
          |--B
            |--he_000.png
            :
            |--he_099.png

  Note:

  The image's name of category dataset must add category label in the end of name , which split of '_'. Such as: img_001_0.png, img_002_0.png, img_003_1.png, img_004_1.png. The last number is the label of the dataset.

3. Training

(1) Download the data and split the images in the correct place.

(2) Run the following command to train the model.

• Single GPU

  cd /CCM-stainGAN/
  
  # training by CCM-stainGAN
  python train.py --gpu_ids 0 --end_epoch 100 --batch_size 4 --batch_size_cls 16 --num_classes 9 --decay_epoch 50 --threshold_A 35 --threshold_B 180 --env ccm
  
  # training by CycleGAN
  python train_cycle.py --gpu_ids 0 --end_epoch 100 --batch_size 4 --decay_epoch 50 --env cycle
  
  # training by UTOM
  python train_utom.py --gpu_ids 0 --end_epoch 100 --batch_size 4 --decay_epoch 50 --threshold_A 35 --threshold_B 180 --env utom

• Multi GPUs

  Please change the gpu_ids to adapt your device.

  cd /CCM-stainGAN/
  
  # training by CCM-stainGAN
  python train.py --gpu_ids 0,1,2,3 --end_epoch 100 --batch_size 4 --batch_size_cls 16 --num_classes 9 --decay_epoch 50 --threshold_A 35 --threshold_B 180 --env ccm
  
  # training by CycleGAN
  python train_cycle.py --gpu_ids 0,1,2,3 --end_epoch 100 --batch_size 4 --decay_epoch 50 --env cycle
  
  # training by UTOM
  python train_utom.py --gpu_ids 0,1,2,3 --end_epoch 100 --batch_size 4 --decay_epoch 50 --threshold_A 35 --threshold_B 180 --env utom

• Multi GPUs With DDP

  We also provide the DistributedDataParallel version of multi GPUs script. Please change the gpu_ids to adapt your device.

  cd /CCM-stainGAN/
  
  # training by CCM-stainGAN
  python -m torch.distributed.launch --nproc_per_node=6 --master_port 11111 train_ddp.py --gpu_ids 0,1,2,3,4,5 --end_epoch 100 --batch_size 4 --batch_size_cls 8 --num_classes 9 --decay_epoch 50 --threshold_A 35 --threshold_B 180 --env ccm

The models will be saved in /CCM-stainGAN/output/{env}/.

4. Test

• Run the following command to test the result.

• Please change the pretrained_model_path and gpu_ids to adapt your model and device.

  cd /CCM-stainGAN/
  
  # transform by CCM-stainGAN
  python test.py --gpu_ids 0 --pretrained_model_path epoch100.pth --outf ./output/test 

Citation

If the repo helps you, please consider citing our works:

@article{wang2022resection,
  title={Resection-inspired histopathological diagnosis of cerebral cavernous malformations using quantitative multiphoton microscopy},
  author={Wang, Shu and Li, Yueying and Xu, Yixuan and Song, Shiwei and Lin, Ruolan and Xu, Shuoyu and Huang, Xingxin and Zheng, Limei and Hu, Chengcong and Sun, Xinquan and others},
  journal={Theranostics},
  volume={12},
  number={15},
  pages={6595},
  year={2022},
  publisher={Ivyspring International Publisher}
}

@article{wang2023deep,
  title={A deep learning-based stripe self-correction method for stitched microscopic images},
  author={Wang, Shu and Liu, Xiaoxiang and Li, Yueying and Sun, Xinquan and Li, Qi and She, Yinhua and Xu, Yixuan and Huang, Xingxin and Lin, Ruolan and Kang, Deyong and others},
  journal={Nature Communications},
  volume={14},
  number={1},
  pages={5393},
  year={2023},
  publisher={Nature Publishing Group UK London}
}