FedMix: Mixed Supervised Federated Learning for Medical Image Segmentation.

Area of Science:

• Artificial Intelligence
• Medical Imaging
• Machine Learning

Background:

• Federated learning enables collaborative model training without data sharing.
• Existing medical image segmentation models assume uniform data annotation levels.
• This uniformity assumption limits the application of federated learning in diverse clinical settings.

Purpose of the Study:

• To propose FedMix, a label-agnostic federated learning framework for medical image segmentation.
• To enable joint training using data with varying levels of supervision (pixel, bounding box, image-level labels).
• To relax the unrealistic assumption of similar annotation standards across clients.

Main Methods:

• FedMix integrates diverse labeled data (pixel, bounding box, image-level) for local model updates.
• An adaptive weight assignment procedure dynamically adjusts client contributions during global model aggregation.
• The framework utilizes mixed image labels for robust feature representation.

Main Results:

• FedMix significantly outperforms state-of-the-art methods on multiple public datasets.
• The framework demonstrates superior performance by overcoming single-level supervision constraints.
• Adaptive weight assignment enhances feature discrimination and model performance.

Conclusions:

• FedMix offers a flexible and effective solution for federated medical image segmentation with heterogeneous data.
• The framework is extendable to multi-class segmentation and shows clinical feasibility.
• FedMix advances privacy-preserving machine learning in healthcare by leveraging varied data annotations.