Related Experiment Videos
Federated continual learning for privacy-preserving chest radiograph classification
Anay Sinhal1, Amit Sinhal2, Arpana Sinhal3
1University of Florida, Gainesville, USA.
Scientific Reports
|June 10, 2026
Summary
Federated continual learning (FCL) addresses privacy and changing data in medical AI. DP-FedEPC uses elastic weight consolidation and prototypes to maintain knowledge without raw images, improving model performance.
Area of Science:
- Artificial Intelligence
- Medical Imaging
- Machine Learning
Background:
- Deep learning for chest radiographs faces challenges with data privacy and evolving clinical workflows.
- Federated learning (FL) addresses privacy by avoiding raw image pooling, but struggles with non-static data distributions.
- Federated continual learning (FCL) tackles both issues, but existing methods often require data unsuitable for healthcare settings.
Purpose of the Study:
- To propose a novel Federated Continual Learning (FCL) method, DP-FedEPC, designed for multi-site deep learning on chest radiographs.
- To address privacy regulations and changing clinical workflows without relying on raw image replay or public datasets.
- To maintain model performance and knowledge consolidation across sequential updates in a federated environment.
Main Methods:
- DP-FedEPC combines Differentially Private Stochastic Gradient Descent (DP-SGD) on clients, Elastic Weight Consolidation (EWC), and prototype-based rehearsal.
- EWC prevents model parameters from drifting during sequential task updates.
- Latent prototypes, not raw images, are used to stabilize class representations across evolving tasks.
Main Results:
- The study trained models on CheXpert and validated on MIMIC-CXR, reporting performance metrics including macro-AUROC and per-finding accuracy.
- DP-FedEPC demonstrated effective knowledge retention, minimizing forgetting across task shifts.
- The method achieved competitive performance while adhering to privacy constraints through client-side DP-SGD, with explicit [Formula: see text] values reported for different noise levels.
Conclusions:
- DP-FedEPC offers a practical solution for federated continual learning in medical imaging, balancing privacy, performance, and adaptability.
- The proposed method effectively consolidates knowledge across sequential updates without compromising patient privacy or requiring sensitive data sharing.
- This approach facilitates the development of robust and continuously improving deep learning models for chest radiograph analysis in real-world clinical settings.
