SplitAVG: A Heterogeneity-Aware Federated Deep Learning Method for Medical Imaging.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Medical Imaging

Background:

• Federated learning enables collaborative model training without data sharing.
• Data heterogeneity across institutions often degrades federated learning model performance.

Purpose of the Study:

• To propose SplitAVG, a novel method to overcome performance degradation caused by data heterogeneity in federated learning.
• To develop a heterogeneity-aware federated learning approach that avoids complex tuning.

Main Methods:

• SplitAVG utilizes network splitting and feature map concatenation.
• Encourages federated model training of an unbiased estimator for target data distribution.
• Compared SplitAVG against seven state-of-the-art federated learning methods.

Main Results:

• SplitAVG achieved performance comparable to the baseline across all heterogeneous settings.
• Demonstrated 96.2% accuracy in diabetic retinopathy classification and 110.4% MAE in bone age prediction on heterogeneous data.
• Outperformed other federated learning methods that showed significant performance degradation with increasing heterogeneity.

Conclusions:

• SplitAVG effectively mitigates performance drops due to data distribution variability in federated learning.
• The method is adaptable to various convolutional neural networks (CNNs) and medical imaging tasks.
• SplitAVG offers a robust solution for privacy-preserving collaborative model training in healthcare.