apple2373 / wbc-benchmark

White Blood Cell Classification under Domain Shift

If you find the code useful, please consider citing the following paper:

• https://arxiv.org/abs/2303.01777
• Satoshi Tsutsui, Zhengyang Su, and Bihan Wen. (2023). Benchmarking White Blood Cell Classification Under Domain Shift. IEEE International Conference on Acoustics, Speech, & Signal Processing (ICASSP).

@inproceedings{tsutsui2023wbc,
	author = {Satoshi Tsutsui and Zhengyang Su and Bihan Wen},
	booktitle = {IEEE International Conference on Acoustics, Speech, & Signal Processing (ICASSP)},
	month = {06},
	title = {Benchmarking White Blood Cell Classification Under Domain Shift},
	year = {2023}
}

Dataset Preparation

• RaabinWBC
  • Download and decompress Train, TestA, and TestB into ./data/RaabinWBC/ from https://raabindata.com/free-data/.
• LISC
  • Download and decompress LISC into ./data/ from https://drive.google.com/file/d/1gknVrSs1CRy8PoIh1HXiGu-1ObH3cQ9S, which is provided by https://github.com/nimaadmed/WBC_Feature.
  • then python3 scripts/organize_lisc.py, which will make ./data/LISCCropped
• BCCD
  • Download and decompress BCCD into ./data/ from https://github.com/apple2373/BCCD
  • cd data; git clone https://github.com/apple2373/BCCD; cd ../

Environment

• We used python version 3.8.
• We used CUDA version 11.6
• see requirements.txt.
  • We use pytorch with version 1.12.1.
  • We use torchvision with version 0.13.1.
  • We use timm with version 0.6.11.
• We used miniconda. The following commands should do the job.

   conda create -y --name=wbc-benchmark python=3.8
   conda activate wbc-benchmark
   pip install numpy==1.23.1 pandas==1.5.0 Pillow==9.3.0 tqdm==4.64.1
   conda install pytorch==1.12.1 torchvision==0.13.1 torchaudio==0.12.1 cudatoolkit=11.6 -c pytorch -c conda-forge
   # pip install torch==1.12.1+cu116 torchvision==0.13.1+cu116 torchaudio==0.12.1 --extra-index-url https://download.pytorch.org/whl/cu116 # if conda is too slow...
   pip install timm==0.6.11
  

• The exact enviroment we use is see environment.yml.

Usage

• See python3 main.py --help for details.
• By default, it will train on RaabinWBC dataset, and test on LISC dataset.
• E.g., to train ResNet50 with Group Normalization
  • python3 main.py --backbone resnet50_gn
• E.g., to test on RabbinWBC Test A set.
  • python3 main.py --backbone resnet50_gn --test-img-root ./data/RaabinWBC/TestA
• Trained models can be tested on other datasets by feeding --epochs 0 --resume <path to pretrained model> --test-img-root <path to test set root>.

Models

All models are initialized with ImageNet pretrained weights.

Models used in the paper.

• vgg16
  • VGG16
  • python3 main.py --backbone vgg16
• vgg16_bn
  • VGG16 with Batch Normalization
  • python3 main.py --backbone vgg16_bn
• resnet50
  • ResNet50
  • python3 main.py --backbone resnet50
• resnet50-gn
  • ResNet50 that replaced Batch Normalization with Group Normalization
  • python3 main.py --backbone resnet50_gn
• resnet50-fc
  • ResNet50 + VGG-style fully-connected layers.
  • python3 main.py --backbone resnet50 --addfc 1
• resnet50-freezebn
  • ResNet50 + freeze batch normalization layers with ImageNet pretrained parameters
  • python3 main.py --backbone resnet50 --freezebn 1
• resnet50-freezebn-last16only
  • ResNet50 + freeze batch normalization layers with ImageNet pretrained parameters + fine-tune only the last 16 layers.
  • python3 main.py --backbone resnet50 --freezebn 1 --last16only 1

Models not used in the paper.

• resnet50-gn-mixup
  • ResNet50 that replaced Batch Normalization with Group Normalization
  • python3 main.py --backbone resnet50_gn --mixup 1
• vit_b_16
  • Vision Transformer (ViT) Base 16x16.
  • python3 main.py --backbone vit_b_16

Accuracy (%)

The results are obtained by running the model 10 times with random seeds of 0-9. See train_resnet50_gn.sh for example. We report mean accuracy with 95% confidence interval.

• vgg16
  • testA: 98.75 +- 0.06
  • testB: 64.79 +- 6.05
  • LISC: 74.44 +- 2.72
  • BCCD: 17.32 +- 3.40
• vgg16-bn
  • testA: 98.64 +- 0.18
  • testB: 12.06 +- 5.46
  • LISC: 32.33 +- 6.17
  • BCCD: 19.86 +- 2.23
• resnet50
  • testA: 98.53 +- 0.18
  • testB: 34.55 +- 10.72
  • LISC: 23.11 +- 3.04
  • BCCD: 23.19 +- 7.87
• resnet50-fc
  • testA: 98.46 +- 0.19
  • testB: 35.63 +- 4.51
  • LISC: 27.74 +- 3.44
  • BCCD: 21.34 +- 3.04
• resnet50-freezebn
  • testA: 98.94 +- 0.06
  • testB: 65.38 +- 2.78
  • LISC: 51.48 +- 7.14
  • BCCD: 11.97 +- 2.46
• resnet50-freezebn-last16only
  • testA: 98.67 +- 0.12
  • testB: 61.92 +- 2.11
  • LISC: 74.24 +- 2.46
  • BCCD: 23.33 +- 1.21
• resnet50-gn
  • testA: 98.24 +- 0.14
  • testB: 77.95 +- 5.43
  • LISC: 73.07 +- 2.07
  • BCCD: 49.12 +- 6.94
• resnet50-gn-mixup
  • testA: 98.44 +- 0.09
  • testB: 19.67 +- 4.71
  • LISC: 63.50 +- 2.65
  • BCCD: 53.11 +- 4.82
• vit_b_16
  • testA: 98.33 +- 0.14
  • testB: 85.06 +- 2.74
  • LISC: 69.77 +- 3.09
  • BCCD: 41.00 +- 6.66