Contrastive prototype federated learning against noisy labels in fetal standard plane detection

Area of Science:

• Medical Imaging
• Machine Learning
• Federated Learning

Background:

• Federated learning (FL) in medical imaging faces challenges with noisy labels and data heterogeneity.
• Accurate fetal standard plane detection is crucial for prenatal diagnostics.

Purpose of the Study:

• To improve federated learning for ultrasound fetal standard plane detection.
• To address challenges of noisy labels and data size variability in decentralized clients.
• To propose a federated denoising framework using prototypes to refine labels and enhance predictions while preserving privacy.

Main Methods:

• Applied contrastive learning (SimCLR) to the largest dataset for robust embeddings.
• Utilized embeddings and a k-nearest neighbors strategy for noisy label refinement.
• Shared image prototypes and a trained backbone with smaller clients to guide FL.
• Implemented an ensemble strategy with majority voting to optimize label refinement and minimize data discard.

Main Results:

• The proposed framework demonstrated superior performance over traditional FL methods.
• Achieved the highest mean F1-score across all fetal standard planes.
• Showcased improved accuracy in standard plane detection.

Conclusions:

• The federated denoising strategy effectively enhances fetal standard plane detection.
• High-quality prototypes enable robust performance despite noisy and heterogeneous data.
• The approach ensures privacy preservation in federated learning settings.