Help a hematologist challenge

We placed third at the Help a hematologist out challenge. Here is the solution.

Training Cycle-GAN for domain adaptation

We used the Cycle-GAN model to train a generator for generating Mat_19/Ace_20 images from WBC1 images and vice versa.

python train_mean_std.py --name mat_ace_wbc_mean_std --model cycle_gan \
--pool_size 50 --no_dropout --batch_size 16 --netG resnet_9blocks_noTanh

It expects the following in the root directory of the repo:

• metadata.csv file
• Datasets/Acevedo_20, Datasets/Matek_19, Datasets/WBC1 datasets
• Datasets/Mean_image.pickle and Datasets/Std_image.pickle which are the channel-wise mean and standard deviation of the Mat_19 and Ace_20 images

It saves the log and model files under checkpoints/mat_ace_wbc_mean_std directory.

It also saves example generated images, Mat_19/Ace_20 <----> WBC1 after each epoch under figures_mean_std

Using trained Cycle-GAN to generate WBC1-like images for the Mat_19/Ace_20 images

python test_mean_std.py --name mat_ace_wbc_mean_std --model cycle_gan \
--no_dropout --epoch 25 --results_dir Datasets/MAT_ACE_AS_WBC_MEAN_STD \
--netG resnet_9blocks_noTanh

It uses the saved model at epoch 25 (you can change it to other epochs) to generate WBC1-like images for Mat_19/Ace_20 dataset and saves them at Datasets/MAT_ACE_AS_WBC_MEAN_STD

The generated images (Datasets/MAT_ACE_AS_WBC_MEAN_STD) are then used to train a resnet18 classifier model. The trained model is used for making predictions on the dev phase (WBC1) and test phase (WBC2) datasets, evaluated on the challenge website.

Data Preprocessing Details

All the images are RGB. They were first center-cropped to 25 x 25 micrometers to keep the area of the background of the cell the same in all images. Then they were resized to 224 x 224 pixels to fit the input sizes of the GAN and the Resnet18. A mean and standard deviation image was calculated from the resized images from the training+validation dataset (ACE-20 and MAT-19) for all the 3 channels. All images were then channelwise standardised with respect to the respective mean and standard deviation images before training the GAN.

For training the classifier (Resnet18) , the images were geometrically augmented by Vertical and Horizontal Flipping. The RGB Intensity augmentation was done using fancy_pca (https://github.com/pixelatedbrian/fortnight-furniture/blob/master/src/fancy_pca.py). The value for alpha_std used was 0.1 as proposed by the authors of the paper --> http://papers.nips.cc/paper/4824-imagenet-classification-with-deep-convolutional-neural-networks.pdf

Training and making inference using resnet18

To train a classifier resnet18, run the following:

python train_resnet.py &> train.log

It will train the resnet18 model on the fake WBC dataset generated from Mat_19/Ace_20 dataset by the Cycle-GAN

model files and results will be saved under models/resnet_train and results/resnet_train folder.

The csv file to be submitted in the dev phase of the data challenge is at results/resnet_train/submission.csv

Then, the trained model can be used to make prediction on WBC2 (test phase) datasets that is expected to be downloaded and saved at Datasets/WBC2/DATA-TEST

python test_resnet.py &> test.log

The results will be saved under results/resnet_test/submission.csv file which can be uploaded to test phase in the data challenge

The Team

BLAMAD: our team name is basically the first letter of the first names of all team members.

Team Members: Bashir Kazimi, Lea Gabele, Ankita Negi, Martin Brenzke, Arnab Majumdar, and Dawit Hailu

Instructions using the code

The code is copied from here and adapted.

For detailed instruction on on its use, please go to their repo linked above.

Thanks to the original authors of Cycle-GAN.

Citation

Citations for the original Cycle-GAN publications:

@inproceedings{CycleGAN2017,
  title={Unpaired Image-to-Image Translation using Cycle-Consistent Adversarial Networks},
  author={Zhu, Jun-Yan and Park, Taesung and Isola, Phillip and Efros, Alexei A},
  booktitle={Computer Vision (ICCV), 2017 IEEE International Conference on},
  year={2017}
}


@inproceedings{isola2017image,
  title={Image-to-Image Translation with Conditional Adversarial Networks},
  author={Isola, Phillip and Zhu, Jun-Yan and Zhou, Tinghui and Efros, Alexei A},
  booktitle={Computer Vision and Pattern Recognition (CVPR), 2017 IEEE Conference on},
  year={2017}
}

Citations for fancy_pca:

@inproceedings{NIPS2012_c399862d,
 title = {ImageNet Classification with Deep Convolutional Neural Networks},
 author = {Krizhevsky, Alex and Sutskever, Ilya and Hinton, Geoffrey E},
 booktitle = {Advances in Neural Information Processing Systems},
 year = {2012}
}

Acknowledgments

CycleGAN code copied from here and adapted. The fancy_pca implementation was taken from here.