Federated Active Learning for Multicenter Collaborative Disease Diagnosis.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Collaborative Diagnosis

Background:

• Deep learning-based computer-aided diagnosis is crucial in medical imaging.
• Multicenter collaborative diagnosis is a growing trend.
• Centralized learning faces challenges in data annotation, privacy, and generalization.

Purpose of the Study:

• To propose novel federated active learning methods for multicenter collaborative disease diagnosis.
• To address the challenges of large-scale annotation and privacy concerns in medical AI.
• To enhance both data and client efficiency in federated learning settings.

Main Methods:

• Labeling Efficient Federated Active Learning (LEFAL) uses a task-agnostic hybrid sampling strategy for data uncertainty and diversity.
• Training Efficient Federated Active Learning (TEFAL) employs a discriminator to evaluate client informativeness.
• Both methods are designed for multicenter collaborative diagnosis scenarios.

Main Results:

• LEFAL achieved 95% performance on a gastrointestinal disease segmentation task using only 65% of labeled data.
• TEFAL reached 0.90 accuracy and 0.95 F1-score on a COVID-19 classification task within 50 iterations.
• Experimental results show superior performance over state-of-the-art methods in both segmentation and classification tasks.

Conclusions:

• The proposed LEFAL and TEFAL methods effectively address limitations of traditional federated learning in multicenter medical diagnosis.
• These federated active learning approaches significantly improve efficiency and performance in collaborative medical image analysis.
• The study demonstrates the potential of federated active learning for robust and privacy-preserving multicenter disease diagnosis.