Federated learning-based CT liver tumor detection using a teacher‒student SANet with semisupervised learning

Area of Science:

• Medical Imaging Analysis
• Machine Learning in Healthcare
• Radiology AI

Background:

• Computed tomography (CT) scan analysis for liver tumors is crucial but labor-intensive.
• Training effective machine learning models requires extensive expert annotations.
• Current methods face challenges in efficiency and data requirements.

Purpose of the Study:

• To develop an innovative approach for improving CT-based liver tumor detection.
• To significantly reduce the labor and time costs associated with training detection models.
• To enhance model accuracy while minimizing reliance on large annotated datasets.

Main Methods:

• Utilized federated learning for privacy-preserving collaborative model training across institutions.
• Implemented a teacher-student framework with an enhanced slice-aware network (SANet).
• Employed semisupervised learning (SSL) techniques to leverage pseudolabels from a teacher model.

Main Results:

• The proposed method achieved a model accuracy of 83% in simulation experiments.
• Demonstrated improvement over original locally trained models.
• Successfully reduced dependence on extensively annotated datasets.

Conclusions:

• The combined approach of federated learning, teacher-student SANet, and SSL offers a promising solution for CT liver tumor detection.
• The method enhances detection accuracy while substantially lowering labor and time costs.
• Achieved 83% model accuracy, representing a significant advancement over previous techniques.