Advancing COVID-19 Diagnosis with Privacy-Preserving Collaboration in Artificial Intelligence

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Digital Health

Background:

• Concerns about data security and trustworthiness hinder the development of AI models for COVID-19 diagnosis.
• Collecting large-scale, representative medical data for AI training is challenging due to privacy issues.

Approach:

• Introduced the Unified CT-COVID AI Diagnostic Initiative (UCADI) utilizing a federated learning (FL) framework.
• Developed a distributed AI model trained across multiple institutions without direct data sharing.
• Evaluated the FL model on diverse datasets, including hold-out and contrast-enhanced CT scans.

Key Points:

• The FL model demonstrated superior performance over local models, achieving high sensitivity and specificity in China and the UK.
• AI model performance was comparable to that of professional radiologists.
• Visual explanations were provided for AI decisions, and trade-offs between performance and communication costs were analyzed.

Conclusions:

• Federated learning offers a viable solution for privacy-preserving AI in healthcare, enabling robust COVID-19 diagnosis.
• The UCADI initiative advances the application of federated learning for AI in digital health.
• This approach addresses data privacy concerns while facilitating the development of accurate diagnostic AI models.