Abstract

Low-dose computed tomography (LDCT) enables imaging with minimal radiation exposure but results in noisy outputs. Deep learning offers denoising solutions requiring large, multi-center datasets, but privacy and heterogeneity pose challenges. We propose FedFDD, a personalized federated learning approach that decomposes images via discrete cosine transform: high-frequency components are aggregated across clients while preserving local low-frequency structures. Evaluations on organ-specific LDCT datasets demonstrate that FedFDD outperforms state-of-the-art federated and centralized models, especially in non-IID scenarios and challenging denoising cases.

Citation

Chen, X., Li, Z., Xu, Z., Ouyang, C., & Qin, C. (2024). FedFDD: Federated Learning with Frequency Domain Decomposition for Low-Dose CT Denoising. In Medical Imaging with Deep Learning (MIDL 2024), 238:1–16.

@inproceedings{chen2024fedfdd,
  author    = {Xuhang Chen and Zeju Li and Zikun Xu and Cheng Ouyang and Chen Qin},
  title     = {FedFDD: Federated Learning with Frequency Domain Decomposition for Low-Dose CT Denoising},
  booktitle = {Medical Imaging with Deep Learning (MIDL)},
  year      = {2024},
  volume    = {238},
  pages     = {1--16},
  keywords  = {federated learning, low-dose CT, frequency domain decomposition, medical imaging}
}