PointTransformerFL

Official PyTorch implementation for the following paper:

Point Transformer with Federated Learning for Predicting Breast Cancer HER2 Status from Hematoxylin and Eosin-Stained Whole Slide Images

Bao Li, Zhenyu Liu, Lizhi Shao, Bensheng Qiu, Hong Bu, Jie Tian

[arXiv]

🔥 News

• 2023.12: 🎉 PointTransformerFL is accepted by AAAI 2024!

PointTransformerDDA+

The overall framework of PointTransformerDDA+.

Environment Settings

conda create --name point python=3.8
conda activate point
# pytorch 1.12.0 with cuda 11.3
conda install pytorch==1.12.0 torchvision==0.13.0 torchaudio==0.12.0 cudatoolkit=11.3 -c pytorch
# compared graph model are implemented by DGL
conda install -c dglteam/label/cu113 dgl=1.0.2
pip install -r requirements.txt

Run Experiments

For feature extractor, we use the model trained with nuclei instance segmentation tasks. Please refer the repo NuHTC.

Demo for debug the model. We put some point data from public HEROHE dataset in the HER2 directory

# Training using HEROHE and with external dataset HER2C
python main_her2_FL.py --max_epochs 200 --batch_size 4 --fl_avg FedAvg --exp_code fed_avg_demo --aux 1.0 --csv_path dataset_csv/HEROHE_HER2.csv --data_dir HER2 --ind_name her2c

We recommend to use wandb to visualize the model's training progress. To use wandb, please add the argument --wandb in the command line

python main_her2_FL.py --max_epochs 200 --batch_size 32 --fl_avg FedAvg --exp_code fed_avg_demo --aux 1.0 --fast_sim --csv_path data_csv/HEROHE_HER2.csv --ind_name her2c --wandb

Experiments runed in our study, need the whole data.

# Centralized Training
python main_her2_FL.py --exp_code central_baseline --no_fl
# Base PointTransformer
python main_her2_FL.py --fl_avg FedAvg --exp_code PointTransformer
# PointTransformer with FCS
python main_her2_FL.py --fl_avg FedAvg --exp_code PointTransformer+ --fast_sim
# PointTransformer with DDA
python main_her2_FL.py --fl_avg FedAvg --exp_code PointTransformerDDA --aux 1.0
# PointTransformer with FCS and DDA
python main_her2_FL.py --fl_avg FedAvg --exp_code PointTransformerDDA+ --aux 1.0 --fast_sim

Reference By

qq456cvb/Point-Transformers
mahmoodlab/HistoFL

Cite

If it is helpful for your work, please cite this paper:

@inproceedings{li2024point,
  title={Point Transformer with Federated Learning for Predicting Breast Cancer HER2 Status from Hematoxylin and Eosin-Stained Whole Slide Images},
  author={Li, Bao and Liu, Zhenyu and Shao, Lizhi and Qiu, Bensheng and Bu, Hong and Tian, Jie},
  booktitle={Proceedings of the AAAI Conference on Artificial Intelligence},
  volume={38},
  number={4},
  pages={3000--3008},
  year={2024}
}