Federated Pseudo-Labeling: A Data-Centric, Privacy-Preserving Framework for Medical Image Segmentation

Area of Science:

• Medical Image Analysis
• Artificial Intelligence
• Data Privacy

Background:

• Patient privacy concerns and inconsistent annotations limit medical data sharing for deep learning.
• Current federated learning (FL) methods face challenges with uniform architectures, privacy risks, and communication costs.

Purpose of the Study:

• To introduce DCFed, a data-centric, semi-supervised framework for privacy-preserving medical image segmentation.
• To address limitations of data sharing and conventional FL in medical AI.

Main Methods:

• Utilized pseudo-labeling and uncertainty estimation on public, unannotated datasets.
• Implemented a modified U-Net architecture with residual blocks, ASPP, and CBAM at the client level.
• Developed a data-centric, semi-supervised approach avoiding raw data and parameter sharing.

Main Results:

• DCFed improved performance by up to 8.9% on breast cancer ultrasound and 3.7% on skin cancer dermoscopy datasets compared to local training.
• Outperformed FedAvg and FedNova in multi-client scenarios for both medical imaging tasks.
• Demonstrated superior results over centralized training on local data and parameter-sharing FL.

Conclusions:

• DCFed offers a scalable and privacy-preserving solution for medical image segmentation.
• The framework effectively leverages public data to enhance model generalizability and performance.
• Surpasses existing FL and centralized training methods in privacy and efficiency.