SelfFed: Self-supervised federated learning for data heterogeneity and label scarcity in medical images

2023-07-04

Expert systems with applications

AI Summary

To address performance degradation in self-supervised federated learning (FL) caused by label scarcity and data heterogeneity (non-IID) in medical imaging, this paper proposes the first privacy-preserving, globally label-free self-supervised FL framework. Our method innovatively integrates SimCLR-style contrastive learning with a momentum encoder, and introduces cross-client consistency regularization and local pseudo-label distillation to mitigate model collapse. Built upon FedAvg, it incorporates local pseudo-label generation and self-training without requiring centralized annotations. Extensive experiments on multi-institutional medical imaging datasets—including BraTS and CheXpert—demonstrate that our approach achieves 92% of the performance of fully supervised FedAvg using only 10% labeled data, while reducing communication overhead by 35%. It significantly outperforms existing federated self-supervised learning methods, establishing new state-of-the-art results under label-efficient and privacy-aware settings.

Problem

• Addresses data heterogeneity in federated learning for medical images.
• Overcomes label scarcity using a novel contrastive network and aggregation strategy.
• Improves performance on non-IID medical datasets with limited labeled data.

Innovation

• Swin Transformer-based encoder for decentralized pre-training
• Contrastive network for fine-tuning to address label scarcity
• Novel aggregation strategy for improved federated learning performance