FedHAC: Towards Robust Federated Multi-Lesion Segmentation With Heterogeneous Annotation Completeness

Area of Science:

• Artificial Intelligence
• Medical Imaging
• Computer Vision

Background:

• Federated learning (FL) enables collaborative medical image segmentation while protecting patient privacy.
• Existing FL methods often overlook annotation completeness heterogeneity, a common issue in clinical settings.
• This oversight hinders the practical deployment of FL for medical image analysis.

Purpose of the Study:

• To address the challenge of annotation incompleteness in federated medical image segmentation.
• To propose a novel framework, FedHAC, designed for robustness against incomplete annotations.
• To enhance the performance and reliability of collaborative medical image segmentation models.

Main Methods:

• FedHAC employs three modules: Global Class Prototype Alignment (GCPA), Annotation Completeness-Aware Aggregation (ACAA), and GMM-driven Progressive Correction (GPC).
• GCPA establishes a robust initial model using proximal regularization and prototype alignment.
• ACAA assesses annotation completeness per client, prioritizing those with higher quality data.
• GPC utilizes Gaussian Mixture Models (GMM) to classify clients as 'noisy' or 'clean' for progressive error correction.

Main Results:

• FedHAC demonstrated superior performance compared to state-of-the-art methods in medical image segmentation.
• The framework effectively handles various levels of annotation incompleteness.
• Ablation studies confirmed the contribution of each module within FedHAC.

Conclusions:

• FedHAC offers a robust solution for federated medical image segmentation, specifically addressing annotation incompleteness.
• The proposed method significantly improves segmentation accuracy in the presence of incomplete data.
• This work paves the way for more reliable FL applications in clinical medical image analysis.