Relay learning: a physically secure framework for clinical multi-site deep learning.

Area of Science:

• Medical Artificial Intelligence
• Deep Learning in Healthcare
• Data Security and Privacy

Background:

• Multi-site clinical data is crucial for robust deep learning models in healthcare.
• Data sharing is hindered by significant security and privacy concerns.
• Current multi-site learning methods rely heavily on network security infrastructure.

Purpose of the Study:

• To introduce Relay Learning, a secure deep learning framework for multi-site clinical data analysis.
• To address data security and privacy challenges in collaborative medical AI.
• To leverage multi-site big data while ensuring local data isolation.

Main Methods:

• Development of the Relay Learning framework, which physically isolates clinical data.
• Demonstration of efficacy across three distinct medical tasks: retinal fundus segmentation, mediastinum tumor diagnosis, and brain midline localization.
• Evaluation through multi-site and external validation using 41,038 medical images from 21 hosts.

Main Results:

• Significant performance improvements observed across all three medical tasks using Relay Learning.
• Average performance increases of 44.4% (retinal segmentation), 24.2% (tumor diagnosis), and 36.7% (brain localization).
• Relay Learning outperformed central learning on external test sets, maintaining local data sovereignty.

Conclusions:

• Relay Learning offers a secure and effective solution for multi-site deep learning in healthcare.
• The framework enhances collaboration without cross-site network data transfer.
• Relay Learning has the potential to transform clinical multi-site collaboration and advance healthcare.