[A low- dose CT reconstruction algorithm across different scanners based on federated feature learning].

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer Vision

Context:

• Deep learning models for CT reconstruction often struggle with generalization across different scanners and protocols.
• Data privacy concerns limit the sharing of sensitive patient data for model training.
• Federated learning offers a privacy-preserving approach but can face challenges in model generalization.

Purpose:

• To propose a novel federated learning framework, Federated Feature Learning for CT (FedCT), for low-dose CT reconstruction.
• To improve the generalization of deep learning models across multiple CT scanners and protocols.
• To enhance CT image reconstruction performance while ensuring data privacy.

Summary:

• The FedCT framework utilizes an inverse Radon transform-based local network model within a federated learning structure.
• A projection-domain specific learning strategy preserves geometric information, and federated feature learning with conditional parameters enhances image-domain generalization.
• Experiments demonstrate FedCT's superior performance in peak signal-to-noise ratio (PSNR), structural similarity index measure (SSIM), and root-mean-square error (RMSE) compared to existing federated learning methods.

Impact:

• FedCT provides an effective solution for collaborative CT reconstruction model development.
• The framework enhances model generalization and improves global reconstruction performance.
• Data privacy is maintained throughout the collaborative training process.