FedGSCA: Medical Federated Learning With Global Sample Selector and Client Adaptive Adjuster Under Label Noise

Area of Science:

• Artificial Intelligence
• Medical Imaging
• Machine Learning

Background:

• Federated Learning (FL) enables collaborative medical image classification while protecting data privacy.
• Label noise from data variability causes instability and performance degradation in FL models.
• Current FL methods struggle with heterogeneous noise and imbalanced medical data.

Purpose of the Study:

• To propose FedGSCA, a novel framework to enhance robustness in noisy medical FL.
• To address noise heterogeneity and improve global model stability.
• To mitigate the impact of noisy data and prevent overfitting in local training.

Main Methods:

• FedGSCA utilizes a Global Sample Selector to aggregate noise knowledge across clients.
• A Client Adaptive Adjustment (CAA) mechanism is developed, incorporating adaptive pseudo-labeling and Robust Credal Labeling Loss.
• CAA dynamically adjusts to class distributions, includes minority samples, and manages noisy labels by considering multiple plausible labels.

Main Results:

• FedGSCA was evaluated on real-world colon slides and synthetic medical datasets under diverse noise conditions.
• The framework demonstrated superior performance compared to state-of-the-art methods, particularly in extreme and heterogeneous noise scenarios.
• FedGSCA significantly improved model stability and handling of complex noise.

Conclusions:

• FedGSCA offers a robust solution for noisy medical federated learning.
• The proposed framework effectively handles noise heterogeneity and data imbalance.
• FedGSCA is well-suited for real-world medical FL applications requiring high accuracy and stability.