FedMRI

Specificity-Preserving Federated Learning for MR Image Reconstruction

Authors: Chun-Mei Feng, Yunlu Yan, Shanshan Wang, Yong Xu, and Ling Shao, Huazhu Fu,

⚡ Dependencies

numpy==1.18.5
scikit_image==0.16.2
torchvision==0.8.1
torch==1.7.0
runstats==1.8.0
pytorch_lightning==1.0.6
h5py==2.10.0
PyYAML==5.4

⚡ Overview

Motivation

Classical FL algorithm for MR image reconstruction: (a) average all the local client models to obtain a general global model, or (b) repeatedly align the latent features between the source and target clients~\cite{guo2021multi}. In contrast, we propose a specificity-preserving mechanism (c) to consider both generalized shared information'' as well as client-specific properties'' in both the frequency and image spaces.

Framework Overview

Overview of the FedMRI framework. Instead of averaging all the local client models, a globally shared encoder is used to obtain a generalized representation, and a client-specific decoder is used to explore unique domain-specific information. We apply the weighted contrastive regularization to better pull the positive pairs together and push the negative ones towards the anchor.

Baselines

Transfer-Site: where the model is transferred across different sites in a random order.

SingleSet: in which each client is trained using their local data without FL;

FedAvg: https://github.com/vaseline555/Federated-Averaging-PyTorch;

FL-MRCM: https://github.com/guopengf/FL-MRCM;

GD-GD: https://github.com/ki-ljl/FedPer;

LG-FedAvg: https://github.com/pliang279/LG-FedAvg?utm_source=catalyzex.com;

FedBN: https://github.com/med-air/FedBN?utm_source=catalyzex.com;

FedProx: https://github.com/litian96/FedProx?utm_source=catalyzex.com;

Qualitative Results

T-SNE visualizations of latent features from four datasets, where (a-d) show the distributions of SingleSet, FedAvg, FedMRI without Lcon, and our entire FedMRI algorithm, respectively.

⚡ Data Prepare

Download data from the link fastMRI:https://fastmri.org/dataset/,

BraTs: https://www.med.upenn.edu/sbia/brats2018/data.html,

SMS and uMR will be released after excluding patient personal information.

[Training code --> FedMRI]

git clone https://github.com/chunmeifeng/FedMRI.git

⚡ Train

single gpu train "python train.py"

python train.py

multi gpu train "python train_multi_gpu.py"

python train_multi_gpu.py

⚡ Contcat

Any problem please feel free to contact me: strawberry.feng0304@gmail.com

⚡ Citation

@article{feng2021specificity,
  title={Specificity-preserving federated learning for mr image reconstruction},
  author={Feng, Chun-Mei and Yan, Yunlu and Wang, Shanshan and Xu, Yong and Shao, Ling and Fu, Huazhu },
  journal={IEEE Transactions on Medical Imaging},
  year={2022}
}