Pseudo-Global Based Sequential Contribution Estimation for Federated Semi-Supervised Medical Image Segmentation

Area of Science:

• Artificial Intelligence
• Medical Imaging
• Machine Learning

Background:

• Federated semi-supervised learning (FSSL) is crucial for medical image segmentation due to limited labeled data.
• Existing FSSL methods often underutilize unsupervised clients, focusing instead on fully supervised ones.

Purpose of the Study:

• To propose a novel FSSL method, FedPSC, that effectively leverages unsupervised clients for medical image segmentation.
• To enhance the contribution estimation of all clients, including unsupervised ones, in federated learning settings.

Main Methods:

• FedPSC estimates client performance contributions using validation performance differences.
• A pseudo-global model is constructed based on estimated contributions.
• Gradient contribution is estimated using the pseudo-global model and exclusion effect, establishing a relationship between contributions.
• A gradient direction exponential moving average method integrates global and local knowledge.

Main Results:

• FedPSC demonstrates effectiveness in medical image segmentation tasks.
• The method successfully leverages information from unsupervised clients.
• Experimental results on two datasets validate the proposed approach.

Conclusions:

• FedPSC offers an effective strategy for improving federated semi-supervised medical image segmentation.
• The method enhances the utilization of diverse client data, including unlabeled datasets.
• Further analysis provides deeper insights into the FedPSC framework.