One-shot Federated Learning on Medical Data using Knowledge Distillation with Image Synthesis and Client Model...

Area of Science:

• Artificial Intelligence
• Machine Learning
• Medical Imaging

Background:

• One-shot federated learning (FL) is valuable when multiple communication rounds are impractical.
• Training robust global models with FL on medical data is challenging due to less discriminative features and overfitting risks.

Purpose of the Study:

• To introduce FedISCA, a novel one-shot FL framework using image synthesis and client adaptation with knowledge distillation.
• To enhance global model training robustness and alleviate data privacy concerns in medical AI.

Main Methods:

• FedISCA generates diverse synthetic images to prevent overfitting and facilitate robust training via knowledge distillation (KD).
• Noise-adapted client models are designed to mitigate domain disparity during synthesis by updating batch normalization statistics.
• The global model is trained iteratively using KD with both original and noise-adapted client models and synthetic images.

Main Results:

• FedISCA demonstrated superior accuracy compared to existing methods on multiple medical image classification datasets.
• The framework effectively alleviates data privacy concerns while enabling robust global model training.

Conclusions:

• FedISCA offers an effective solution for one-shot federated learning in medical imaging, particularly when data is limited or privacy is a concern.
• The proposed image synthesis and client adaptation strategy significantly improves model performance and robustness.