Advancing COVID-19 diagnosis with privacy-preserving collaboration in artificial intelligence

Area of Science:

• Digital Health
• Medical Imaging Analysis
• Artificial Intelligence in Healthcare

Background:

• Artificial intelligence (AI) offers potential for efficient COVID-19 diagnosis via medical imaging.
• Data privacy and security concerns hinder the collection of large, representative datasets for AI model training.
• This limitation poses a significant challenge for developing generalized AI diagnostic models for clinical use.

Purpose of the Study:

• To introduce the Unified CT-COVID AI Diagnostic Initiative (UCADI) utilizing a federated learning framework.
• To enable distributed AI model training and independent execution at host institutions without direct data sharing.
• To address challenges in training generalized AI models for COVID-19 diagnosis while preserving patient privacy.

Main Methods:

• Implemented a federated learning framework for distributed training of an AI diagnostic model.
• Collected 9,573 chest computed tomography (CT) scans from 3,336 patients across 23 hospitals in China and the UK.
• Evaluated model performance against local models, professional radiologists, hold-out data, and heterogeneous data.

Main Results:

• The federated learning model significantly outperformed all local models.
• Achieved high test sensitivity (0.973) and specificity (0.951) in China, and (0.730/0.942) in the UK.
• Demonstrated comparable performance to expert radiologists and robustness on diverse datasets.

Conclusions:

• Federated learning offers a viable solution for privacy-preserving AI in medical diagnostics.
• The UCADI initiative successfully trained a generalized AI model for COVID-19 diagnosis using CT scans.
• This approach advances the potential of federated learning for secure and effective digital health applications.