Medical federated learning with joint graph purification for noisy label learning

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Machine Learning

Background:

• Federated Learning (FL) enables collaborative model training while preserving data privacy, crucial for sensitive medical data.
• Label noise, arising from observer variability, is an inherent challenge in medical dataset preparation.
• FL exacerbates label noise issues due to data inaccessibility and potential noise heterogeneity across clients.

Purpose of the Study:

• To develop a robust federated learning framework addressing label noise in medical imaging.
• To enhance diagnostic model performance despite inherent data imperfections and privacy constraints.

Main Methods:

• Proposed FedGP framework with client-side noisy graph purification for reliable pseudo-label generation.
• Implemented a graph-guided negative ensemble loss for robust supervision against label noise.
• Introduced server-side global centroid aggregation for collaborative optimization and robust global knowledge integration.

Main Results:

• FedGP significantly outperforms existing methods in medical FL settings with homogeneous, heterogeneous, and real-world label noise.
• Demonstrated superior performance on endoscopic and pathological image datasets.
• Achieved substantial improvements over state-of-the-art denoising and noisy FL techniques.

Conclusions:

• The FedGP framework effectively mitigates label noise challenges in federated learning for medical imaging.
• Joint graph purification and global aggregation provide a robust solution for privacy-preserving, accurate medical AI.
• The proposed method offers a significant advancement for real-world applications of FL in healthcare.