Personalized federated learning for abdominal multi-organ segmentation based on frequency domain aggregation.

Area of Science:

• Medical Imaging Analysis
• Artificial Intelligence in Healthcare
• Machine Learning for Diagnostics

Background:

• Deep learning (DL) for medical imaging necessitates large, sensitive datasets, posing privacy and annotation challenges.
• Federated learning (FL) allows collaborative model training without direct data sharing, addressing privacy concerns.
• Existing FL methods struggle with statistical heterogeneity across diverse medical datasets.

Purpose of the Study:

• To introduce a novel Personalized Federated Learning Framework (PAF-Fed) for abdominal multi-organ segmentation.
• To enhance collaborative model training by selectively sharing parameters and preserving local data distributions.
• To improve the aggregation of knowledge from heterogeneous medical imaging datasets.

Main Methods:

• PAF-Fed selectively shares partial model parameters for inter-client collaboration.
• Local parameters are retained to learn site-specific data distributions.
• Fourier Transform with Self-attention aggregates low-frequency parameters to manage statistical heterogeneity.

Main Results:

• The PAF-Fed framework was evaluated on the CHAOS 2019 MRI dataset and a private CT dataset.
• Achieved an average Dice Similarity Coefficient (DSC) of 72.65% on the CHAOS dataset.
• Achieved an average Dice Similarity Coefficient (DSC) of 85.50% on the private CT dataset.

Conclusions:

• PAF-Fed demonstrates superior performance compared to state-of-the-art federated learning methods.
• The selective parameter sharing and Fourier Transform aggregation effectively handle data heterogeneity.
• The framework offers a promising solution for privacy-preserving medical image segmentation.