Distributed contrastive learning for medical image segmentation.

Area of Science:

• Artificial Intelligence
• Medical Imaging
• Machine Learning

Background:

• Supervised deep learning requires extensive labeled data, which is often scarce in medical imaging across different institutions.
• Federated learning (FL) enables decentralized model training but typically demands fully labeled datasets.
• Self-supervised contrastive learning (CL) leverages unlabeled data but struggles with data diversity limitations in federated settings (FCL).

Purpose of the Study:

• To propose two novel federated self-supervised learning frameworks for volumetric medical image segmentation with limited annotations.
• To address the challenges of data scarcity and diversity in federated contrastive learning for medical applications.
• To develop frameworks balancing high accuracy with efficient communication for different deployment scenarios.

Main Methods:

• Framework 1: Exchanges features during federated contrastive learning (FCL) for enhanced local CL, using global structural matching for feature alignment.
• Framework 2: Introduces FCLOpt, an optimized method reducing communication costs by omitting negative samples.
• Framework 2 also incorporates predictive target network update (PTNU) and distance prediction (DP) to minimize model download and upload communications.

Main Results:

• Both proposed frameworks significantly enhance segmentation and generalization performance on a cardiac MRI dataset.
• Framework 1 demonstrates high accuracy suitable for powerful servers.
• Framework 2 achieves lower communication costs, making it viable for mobile devices.

Conclusions:

• The developed federated self-supervised learning frameworks effectively overcome limitations of data scarcity and diversity in medical image segmentation.
• The two distinct frameworks provide flexible solutions for different computational and communication constraints.
• These approaches offer substantial improvements over existing state-of-the-art techniques for federated medical image analysis.