Label-Efficient Self-Supervised Federated Learning for Tackling Data Heterogeneity in Medical Imaging

Area of Science:

• Artificial Intelligence
• Medical Imaging
• Computer Science

Background:

• Deep learning models require large medical datasets, but privacy concerns limit data sharing between institutions.
• Federated learning (FL) offers privacy-preserving collaborative training but struggles with data heterogeneity and limited labeled data.

Purpose of the Study:

• To develop a robust and label-efficient self-supervised FL framework for medical image analysis.
• To improve deep learning model performance in privacy-sensitive, multi-institutional settings.

Main Methods:

• A Transformer-based self-supervised pre-training approach using masked image modeling on decentralized datasets.
• Federated learning framework applied to heterogeneous medical imaging datasets (retinal, dermatology, chest X-ray).

Main Results:

• Masked image modeling with Transformers significantly enhances model robustness against data heterogeneity.
• Achieved accuracy improvements of 5.06%, 1.53%, and 4.58% on retinal, dermatology, and chest X-ray classification, respectively, under severe heterogeneity.
• Outperforms existing FL algorithms in generalization to out-of-distribution data and fine-tuning with limited labels.

Conclusions:

• The proposed self-supervised FL framework effectively addresses data heterogeneity and label scarcity in medical image analysis.
• Transformer-based masked image modeling is a powerful tool for robust representation learning in federated medical AI.
• This approach facilitates more accurate and generalizable medical image analysis while preserving data privacy.