bianchang / bci

BCI: Breast Cancer Immunohistochemical Image Generation through Pyramid Pix2pix

BCI: Breast Cancer Immunohistochemical Image Generation through Pyramid Pix2pix

Project | Arxiv | |

News

• ⚡(2022-10-20): We have released the trained model on BCI and LLVIP datasets.
• ⚡(2022-6-29): We host a competition for breast cancer immunohistochemistry image generation on Grand Challenge！
• ⚡(2022-4-26): We have released BCI dataset and the code of PyramidPix2pix. You can download BCI dataset from the homepage.

---

---

Framework

Setup

1)Envs

• Linux
• Python>=3.6
• CPU or NVIDIA GPU + CUDA CuDNN

Install python packages

git clone https://github.com/bupt-ai-cz/BCI
cd PyramidPix2pix
pip install -r requirements.txt

2)Prepare dataset

• Download BCI dataset from our homepage.

• Combine HE and IHC images.

  Project pix2pix provides a python script to generate pix2pix training data in the form of pairs of images {A,B}, where A and B are two different depictions of the same underlying scene, these can be pairs {HE, IHC}. Then we can learn to translate A(HE images) to B(IHC images).

  Create folder /path/to/data with subfolders A and B. A and B should each have their own subfolders train, val, test, etc. In /path/to/data/A/train, put training images in style A. In /path/to/data/B/train, put the corresponding images in style B. Repeat same for other data splits (val, test, etc).

  Corresponding images in a pair {A,B} must be the same size and have the same filename, e.g., /path/to/data/A/train/1.jpg is considered to correspond to /path/to/data/B/train/1.jpg.

  Once the data is formatted this way, call:

  python datasets/combine_A_and_B.py --fold_A /path/to/data/A --fold_B /path/to/data/B --fold_AB /path/to/data
  

  This will combine each pair of images (A,B) into a single image file, ready for training.

• File structure

  PyramidPix2pix
    ├──datasets
         ├── BCI
               ├──train
               |    ├── 00000_train_1+.png
               |    ├── 00001_train_3+.png
               |    └── ...
               └──test
                    ├── 00000_test_1+.png
                    ├── 00001_test_2+.png
                    └── ...
  
  

Train

Train at full resolution(1024*1024):

python train.py --dataroot ./datasets/BCI --gpu_ids 0 --pattern L1_L2_L3_L4

By default, four scales of the pyramid are used for supervision. You can change the option --pattern to use less scales (e.g. --pattern L1_L2_L3).

Train at resolution 512*512 (less GPU memory required):

python train.py --dataroot ./datasets/BCI --preprocess crop --crop_size 512 --gpu_ids 0 --pattern L1_L2_L3_L4

Images are randomly cropped if trained at low resolution.

Test

Test at full resolution(1024*1024):

python test.py --dataroot ./datasets/BCI --gpu_ids 0

Test at resolution 512*512:

python test.py --dataroot ./datasets/BCI --preprocess crop --crop_size 512 --gpu_ids 0

See PyramidPix2pix/options for more train and test options.

Evaluate

Calculate average PSNR and SSIM.

python evaluate.py --result_path ./results/pyramidpix2pix

Trained models

dataset	device	backbone	PSNR	SSIM	model
BCI	Tesla V100-32GB	resnet_9	21.16	0.477	download
LLVIP	Tesla V100-32GB	resnet_9	12.189	0.279	download

Results

Citation

@InProceedings{Liu_2022_CVPR,
    author    = {Liu, Shengjie and Zhu, Chuang and Xu, Feng and Jia, Xinyu and Shi, Zhongyue and Jin, Mulan},
    title     = {BCI: Breast Cancer Immunohistochemical Image Generation Through Pyramid Pix2pix},
    booktitle = {Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops},
    month     = {June},
    year      = {2022},
    pages     = {1815-1824}
}

Contact

Shengjie Liu ([email protected])

Chuang Zhu ([email protected])

If you have any questions, you can contact us directly.