- ⚡(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 or the-following-link.
- 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
-
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 subfoldersA
andB
.A
andB
should each have their own subfolderstrain
,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 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 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.
Calculate average PSNR and SSIM.
python evaluate.py --result_path ./results/pyramidpix2pix
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 |
@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}
}
Shengjie Liu (shengjie.Liu@bupt.edu.cn)
Chuang Zhu (czhu@bupt.edu.cn)
If you have any questions, you can contact us directly.