[A low-dose CT image restoration method based on central guidance and alternating optimization].

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Federated Learning

Context:

• Low-dose computed tomography (CT) imaging is crucial for reducing radiation exposure.
• Image restoration is essential for maintaining diagnostic quality in low-dose CT.
• Federated learning presents challenges in data heterogeneity across institutions.

Purpose:

• To introduce a novel federated learning framework, Federated Guided Proximal (FedGP), for low-dose CT image restoration.
• To address the issue of CT imaging heterogeneity in multi-institutional settings.
• To enhance the generalization capability of CT imaging models across diverse datasets.

Summary:

• FedGP utilizes an alternating optimization strategy where institutions serve as central servers sequentially.
• An institution-modulated CT image restoration network is employed for client-side local training.
• Central guidance leverages local labeled data to steer client network training, improving model generalization.

Impact:

• FedGP achieved superior performance in low-dose and sparse-view CT image restoration, evidenced by higher PSNR and SSIM, and lower RMSE.
• The framework demonstrated robustness and adaptability to data heterogeneity, outperforming models without central guidance.
• FedGP offers a flexible federated learning solution for multi-institutional CT imaging, improving model generalization under varied imaging configurations.