Accounting for data variability in multi-institutional distributed deep learning for medical imaging.

Area of Science:

• Medical imaging analysis
• Artificial intelligence in healthcare
• Distributed computing

Background:

• Sharing patient data for deep learning is hindered by regulatory and technical issues.
• Distributed learning, sharing model weights instead of data, is a viable alternative.
• Cyclical weight transfer (CWT) is effective for homogeneous medical imaging data.

Purpose of the Study:

• Optimize CWT to handle variability in training sample sizes and label distributions across institutions.
• Improve the generalizability of deep learning models in distributed medical imaging settings.

Main Methods:

• Implemented optimizations including proportional local training iterations, cyclical learning rate, locally weighted minibatch sampling, and cyclically weighted loss.
• Evaluated optimized CWT on simulated distributed datasets for diabetic retinopathy detection and chest radiograph classification.

Main Results:

• Proportional local training iterations improved accuracy with sample size variability, reaching 98.6% of central hosting performance.
• Locally weighted minibatch sampling and cyclically weighted loss improved accuracy with label distribution variability, reaching 98.6% and 99.1% respectively.
• Optimized CWT significantly reduced performance gaps compared to non-optimized CWT and central hosting.

Conclusions:

• Optimizations enhance CWT's ability to manage data heterogeneity in distributed medical imaging.
• This study is the first to address sample size and label distribution variability in distributed deep learning for medical imaging.
• Further research is needed to address other real-world data variability challenges.