Hypernetwork-Based Physics-Driven Personalized Federated Learning for CT Imaging.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Radiology

Background:

• Computed Tomography (CT) is crucial for patient anatomy visualization but poses radiation risks.
• Deep learning (DL) for CT reconstruction faces challenges with data centralization, domain shift, and privacy.
• Existing DL methods require large, centralized datasets, limiting personalization and raising privacy concerns.

Purpose of the Study:

• To develop a privacy-preserving, personalized federated learning method for CT imaging.
• To address domain shift and data scarcity issues in DL-based CT reconstruction.
• To improve CT imaging quality through institution-specific adaptations without data sharing.

Main Methods:

• Proposed HyperFed: a hypernetwork-based, physics-driven personalized federated learning framework.
• Utilizes institution-specific physics-driven hypernetworks for local adaptation.
• Employs a global-sharing imaging network to learn invariant features across domains.
• Hypernetworks condition the global network using hyperparameters from specific physical scanning protocols.

Main Results:

• HyperFed achieved competitive performance compared to state-of-the-art methods.
• Demonstrated effective personalized local CT reconstruction.
• Successfully mitigated domain shift and privacy concerns associated with data centralization.
• Validated the approach through experimental evaluations.

Conclusions:

• HyperFed offers a promising direction for enhancing CT imaging quality.
• Enables personalized CT reconstruction tailored to different institutions or scanners.
• Provides a privacy-preserving alternative to centralized DL training for CT.
• Eliminates the need for direct data sharing, preserving patient confidentiality.