Efficient one-shot federated learning on medical data using knowledge distillation with image synthesis and client...

Area of Science:

• Artificial Intelligence
• Machine Learning
• Medical Imaging

Background:

• One-shot federated learning (FL) is crucial when multiple communication rounds are impractical.
• Existing knowledge distillation (KD) methods in one-shot FL struggle with overfitting and high computational demands.
• Medical image analysis requires efficient and accurate distributed learning models.

Purpose of the Study:

• To develop an efficient and robust one-shot federated learning framework for medical image classification.
• To address overfitting and reduce computational complexity in federated learning scenarios.
• To enhance knowledge transfer from clients to a global model using synthetic data.

Main Methods:

• Proposed a novel one-shot FL framework utilizing mixup for generating pseudo-intermediate samples with diverse structure noise.
• Implemented noise-adapted client models by updating batch normalization statistics to mitigate domain disparity.
• Employed iterative knowledge distillation (KD) between original and noise-adapted client models for global model updates.

Main Results:

• The proposed method significantly enhances training sample diversity, effectively preventing overfitting.
• Reusing synthesized images reduces computational resources and improves overall training efficiency.
• Extensive evaluations on multiple medical image classification datasets confirmed the approach's superiority over existing methods.

Conclusions:

• The novel one-shot FL framework effectively tackles overfitting and computational challenges in medical image classification.
• The use of mixup-generated noisy synthetic images enhances model robustness and training efficiency.
• The noise-adapted client models and KD strategy improve knowledge transfer and overall performance.