Federated Semi-Supervised Medical Image Segmentation via Prototype-Based Pseudo-Labeling and Contrastive Learning

Area of Science:

• Medical Imaging Analysis
• Machine Learning
• Federated Learning

Background:

• Federated learning typically requires fully supervised data, which is often unavailable in clinical settings.
• Many clinical sites lack resources or expertise for data annotation, limiting the use of supervised learning.
• Existing federated learning methods are not well-suited for scenarios with a mix of labeled and unlabeled data across clients.

Purpose of the Study:

• To address the challenge of federated semi-supervised segmentation (FSSS) in realistic clinical scenarios.
• To develop a novel FSSS method that leverages unlabeled data effectively.
• To improve segmentation accuracy in medical imaging tasks with limited annotations.

Main Methods:

• Proposed a novel FSSS method combining prototype-based pseudo-labeling and contrastive learning.
• Utilized a labeled-aggregated model for debiased pseudo-label generation on unlabeled clients.
• Implemented prototypical contrastive learning on unlabeled data to enhance feature discrimination.
• Employed a consistency-aware aggregation strategy for dynamic model weighting.

Main Results:

• Demonstrated consistent effectiveness across COVID-19 X-ray and CT segmentation, and colorectal polyp segmentation.
• The proposed method successfully segments infected regions and polyps with limited labeled data.
• Experimental results validate the efficacy of the FSSS approach in practical medical applications.

Conclusions:

• The developed FSSS method offers a viable solution for medical image segmentation tasks with scarce labeled data.
• The integration of pseudo-labeling and contrastive learning significantly improves segmentation performance in federated settings.
• This work paves the way for more practical and resource-efficient federated learning applications in healthcare.