AI-CT-COVID-19

AI-CT-COVID-19 employs federated learning to integrate chest computed tomography (CT) data across institutions to improve diagnostic sensitivity for COVID-19 while preserving patient data privacy.

Key Features:

• Federated Learning Architecture: A decentralized architecture executes model training at host institutions without sharing individual patient data.
• Multi-institution Data Integration: Integrates CT data from multiple institutions, including three Tongji hospitals and Wuhan Union Hospital, to enhance model generalization.
• Diagnostic Performance Metrics: Initial local model trained on three Tongji hospitals achieved 97.5% sensitivity, tested externally on Wuhan Union Hospital at 72% sensitivity, and the federated model achieved 98% sensitivity across test cases.
• UCADI Framework: Uses the UCADI framework to structure global collaboration for building and validating the federated CT-COVID AI model.

Scientific Applications:

• CT-based COVID-19 diagnosis: Enhances sensitivity of chest CT interpretation for COVID-19 detection across institutional datasets.
• Privacy-preserving multi-center training: Enables collaborative model development without transferring raw patient CT data.
• Cross-site validation: Supports validation and generalization of AI models on external datasets such as Wuhan Union Hospital.

Methodology:

An initial AI CT model was trained on localized data from three Tongji hospitals (97.5% sensitivity) and evaluated on Wuhan Union Hospital data (72% sensitivity), then refined via federated learning under the UCADI framework by integrating external datasets without transferring raw patient data to reach 98% sensitivity.