Model-Heterogeneous Semi-Supervised Federated Learning for Medical Image Segmentation.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer Vision

Background:

• Medical image segmentation is vital for clinical diagnosis but faces challenges with data privacy and annotation costs.
• Current collaborative training methods overlook the laborious nature of obtaining segmentation annotations.
• Balancing annotation cost, segmentation performance, and local model personalization is a key research challenge.

Purpose of the Study:

• To introduce a novel Model-Heterogeneous Semi-Supervised Federated (HSSF) Learning framework.
• To reduce the annotation burden and efficiently utilize unlabeled data in medical image segmentation.
• To ensure personalization between different clinical sites.

Main Methods:

• Developed a Model-Heterogeneous Semi-Supervised Federated (HSSF) Learning framework.
• Proposed Regularity Condensation and Regularity Fusion for selective knowledge transfer and personalization.
• Introduced a Self-Assessment (SA) module for real-time confidence generation and a Reliable Pseudo-Label Generation (RPG) module to create pseudo-labels.

Main Results:

• The HSSF framework demonstrated superior performance compared to other methods in heterogeneous settings.
• The model achieved commendable performance in homogeneous designs, particularly in region-based metrics.
• Evaluated on Skin Lesion and Polyp Lesion datasets, showing significant improvements.

Conclusions:

• The HSSF framework effectively addresses the challenges of annotation cost and data privacy in medical image segmentation.
• The proposed SA and RPG modules enhance the utilization of unlabeled data, reducing manual annotation efforts.
• The framework ensures effective knowledge transfer and personalization across different sites, improving overall segmentation accuracy.