A Federated Semi-Supervised Learning Approach for COVID-19 CT Image Segmentation.

Area of Science:

• Medical image analysis
• Artificial intelligence in healthcare
• Federated learning

Background:

• COVID-19 CT image segmentation is crucial but challenged by data scarcity and privacy concerns.
• Effective utilization of limited labeled and abundant unlabeled data is a key issue.
• Inter-institution collaboration is hindered by data privacy protection requirements.

Purpose of the Study:

• To propose an innovative Federated Semi-Supervised Learning method (FSSLCIS) for COVID-19 CT image segmentation.
• To address the challenges of limited labeled data and data privacy in medical image analysis.
• To improve the accuracy and efficiency of COVID-19 CT segmentation.

Main Methods:

• Implemented a Federated Semi-Supervised Learning approach (FSSLCIS).
• Utilized MixMatch for local semi-supervised learning with data augmentation to generate pseudo-labels.
• Employed MixUp strategy for combining labeled and unlabeled data.
• Applied weighted aggregation on the server-side to handle data distribution imbalance.

Main Results:

• The proposed FSSLCIS method demonstrated exceptional performance on three real-world datasets.
• Achieved significant improvements in COVID-19 CT image segmentation accuracy.
• Successfully ensured data privacy during the federated learning process.

Conclusions:

• FSSLCIS is an effective method for COVID-19 CT image segmentation, especially with limited labeled data.
• The method enhances model learning capacity and segmentation accuracy.
• Federated learning with semi-supervised techniques offers a viable solution for privacy-preserving medical image analysis.