Federated Active Learning Framework for Efficient Annotation Strategy in Skin-Lesion Classification.

Area of Science:

• Artificial Intelligence
• Medical Imaging
• Machine Learning

Background:

• Federated learning (FL) allows collaborative model training without data sharing, crucial for privacy-sensitive medical data.
• Data annotation in medical imaging is labor-intensive and requires expertise, posing a significant challenge for FL.
• Active learning (AL) methods can reduce annotation burden by selecting informative samples.

Purpose of the Study:

• To propose a novel federated active learning (AL) framework for medical image analysis.
• To address the critical issue of intensive data annotation requirements in federated learning scenarios.
• To decrease the amount of annotated data needed while maintaining high model performance and patient privacy.

Main Methods:

• Developed a federated AL framework integrating AL periodically and interactively within the FL process.
• Utilized an ensemble of local and global models from FL for data selection.
• Employed ensemble entropy-based AL as an efficient data-annotation strategy.

Main Results:

• The federated AL framework achieved state-of-the-art performance on a skin-lesion classification task using only 50% of the data.
• Outperformed several state-of-the-art AL methods under FL.
• Demonstrated comparable performance to full-data FL while significantly reducing annotation effort and preserving privacy.

Conclusions:

• The proposed federated AL framework effectively reduces data annotation requirements in medical imaging FL.
• This approach maintains high performance and patient privacy, offering a practical solution for medical AI development.
• This represents a novel application of federated AL to medical images, validated on real-world dermoscopic datasets.