Federated Metadata-Constrained iRadonMAP Framework with Mutual Learning for All-in-One Computed Tomography Imaging

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Radiology

Background:

• Growing use of computed tomography (CT) raises concerns about patient radiation dose.
• Deep learning shows potential for improving low-dose CT image quality but struggles with vendor-specific data and heterogeneity.
• Data centralization is limited by cost and privacy regulations, hindering multicenter studies.

Purpose of the Study:

• To develop a generalizable deep learning method for low-dose CT reconstruction across multiple vendors and imaging conditions.
• To address data heterogeneity and privacy concerns in multicenter CT datasets.
• To enable simultaneous reconstruction of multivendor CT images within a single framework.

Main Methods:

• FedM2CT: A federated metadata-constrained method with mutual learning.
• Includes task-specific iRadonMAP (TS-iRadonMAP) for reconstruction, condition-prompted mutual learning (CPML) for knowledge sharing, and federated metadata learning (FMDL) for mitigating data heterogeneity.
• Employs a metamodel to aggregate parameters and handle data variations.

Main Results:

• FedM2CT demonstrated outstanding qualitative and quantitative results in extensive experiments.
• The method achieved successful all-in-one CT reconstruction for various low-dose tasks, including low-milliampere-second, sparse-view, and limited-angle CT.
• Outperformed existing methods in handling multivendor CT data with different imaging geometries and sampling protocols.

Conclusions:

• FedM2CT offers a robust solution for generalizable low-dose CT reconstruction.
• The proposed framework effectively mitigates data heterogeneity challenges in multicenter CT imaging.
• This approach holds significant potential for advancing safe and effective CT imaging practices across diverse clinical settings.