Federated learning for privacy-preserving multi-center tuberculosis diagnosis using chest imaging data.
Authors
Affiliations (6)
Affiliations (6)
- Department of Information Technology, Pillai HOC College of Engineering and Technology, Rasayani, Maharashtra, India. Electronic address: [email protected].
- Department of Information Technology, Vasantdada Patil Pratishthan's College of Engineering & Visual Arts, Sion, Mumbai, Maharashtra, India. Electronic address: [email protected].
- Department of Biotechnology and Microbiology, Noida International University, Uttar Pradesh, 203201, India. Electronic address: [email protected].
- Department of Information Technology, Tulsiramji Gaikwad Patil College of Engineering and Technology, Nagpur, Maharashtra, 441108, India. Electronic address: [email protected].
- Department of Computer Science and Engineering, CSMSS Chh. Shahu College of Engineering, Maharashtra, India. Electronic address: [email protected].
- Department of Computer Technology, Yeshwantrao Chavan College of Engineering, Nagpur, Maharashtra, India. Electronic address: [email protected].
Abstract
Tuberculosis (TB) remains one of the most serious global health challenges, demanding reliable and early detection methods to reduce transmission and mortality. Chest imaging, particularly chest X-rays, plays a vital role in screening, but its interpretation varies across institutions and requires expert radiologists who may not always be available. This study introduces a federated learning framework for privacy-preserving, multi-center TB diagnosis using chest imaging data. Unlike centralized training, the proposed approach allows multiple healthcare institutions to collaboratively train a shared deep learning model without transferring raw patient data, thus ensuring compliance with privacy regulations such as HIPAA and GDPR. Each participating centre trains a local convolutional neural network (CNN) on its dataset, and only encrypted model parameters are shared for aggregation. The framework incorporates differential privacy, secure aggregation, and encryption to enhance confidentiality while maintaining diagnostic accuracy. Extensive experiments on multi-center datasets, including Shenzhen, Montgomery, and NIH ChestX-ray14, demonstrate that the federated CNN achieved an average accuracy of 94.8 %, sensitivity of 93.5 %, and specificity of 95.2 %, closely matching centralized models while offering superior generalization across heterogeneous data sources. Hybrid architectures combining CNN and transformer layers further improved interpretability and precision. The findings confirm that federated learning effectively balances diagnostic performance and patient privacy, establishing a scalable, secure, and collaborative paradigm for medical imaging analysis. This framework provides a promising foundation for broader applications in multi-institutional disease diagnosis and healthcare data governance.