Federated learning for privacy-preserving multi-center tuberculosis diagnosis using chest imaging data

Area of Science:

• Medical Imaging Analysis
• Artificial Intelligence in Healthcare
• Privacy-Preserving Machine Learning

Background:

• Tuberculosis (TB) poses a significant global health threat, necessitating improved diagnostic tools.
• Chest X-rays are crucial for TB screening, but interpretation is subjective and requires expert radiologists.
• Centralized data analysis for medical imaging raises privacy concerns and regulatory challenges (e.g., HIPAA, GDPR).

Purpose of the Study:

• To develop and evaluate a federated learning framework for multi-center, privacy-preserving tuberculosis diagnosis using chest imaging.
• To enable collaborative model training across institutions without compromising patient data confidentiality.
• To assess the diagnostic performance and generalization capabilities of the federated approach.

Main Methods:

• Implementation of a federated learning framework utilizing convolutional neural networks (CNNs) for TB detection.
• Integration of privacy-enhancing techniques: differential privacy, secure aggregation, and encryption.
• Collaborative training across multiple healthcare institutions using local datasets (Shenzhen, Montgomery, NIH ChestX-ray14).
• Evaluation of hybrid CNN-transformer architectures for enhanced interpretability.

Main Results:

• The federated CNN model achieved high diagnostic accuracy (94.8%), sensitivity (93.5%), and specificity (95.2%).
• Performance closely matched centralized training models while demonstrating superior generalization across diverse datasets.
• Hybrid architectures showed potential for improved interpretability and precision in TB diagnosis.

Conclusions:

• Federated learning offers an effective solution for multi-institutional medical imaging analysis, balancing diagnostic accuracy with robust patient privacy.
• The proposed framework establishes a scalable, secure, and collaborative paradigm for disease diagnosis and healthcare data governance.
• This approach facilitates broader applications in collaborative medical research and public health initiatives.