Federated Transfer Learning for Low-dose PET Denoising: A Pilot Study with Simulated Heterogeneous Data

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Radiology

Background:

• Low-dose Positron Emission Tomography (PET) reduces radiation exposure but results in low signal-to-noise ratio (SNR) reconstructions, impacting diagnostic accuracy.
• Deep learning (DL) excels at PET denoising but requires extensive, privacy-protected data, which is difficult to obtain and share across institutions.
• Existing federated learning (FL) methods struggle with domain shift caused by varying low-dose PET protocols across different healthcare facilities.

Purpose of the Study:

• To develop a novel federated transfer learning (FTL) framework for effective low-dose PET denoising.
• To address the challenges of data heterogeneity and privacy concerns in multi-institutional low-dose PET data.
• To improve the performance of PET denoising across institutions with different low-dose settings.

Main Methods:

• Proposed a federated transfer learning (FTL) framework designed for low-dose PET denoising.
• Utilized heterogeneous low-dose PET data from simulated multi-institutional settings.
• Enabled collaborative model training without direct data aggregation, preserving patient privacy.

Main Results:

• The FTL framework successfully leveraged heterogeneous low-dose PET data from different institutions.
• Demonstrated superior low-dose PET denoising performance compared to previous FL methods.
• Effectively addressed the domain shift issue arising from varied low-dose PET protocols.

Conclusions:

• The proposed FTL framework offers an efficient and privacy-preserving solution for low-dose PET denoising.
• This approach enhances image quality and diagnostic utility of low-dose PET scans.
• FTL is a promising technique for advancing AI applications in multi-institutional medical imaging research.