Related Experiment Videos
FedIHRAS: A Privacy-Preserving Federated Learning Framework for Multi-Institutional Collaborative Radiological
André Luiz Marques Serrano1, Gabriel Arquelau Pimenta Rodrigues1, Guilherme Dantas Bispo1
1Department of Electrical Engineering, University of Brasilia, Brasília 70910-900, Brazil.
Biomedicines
|March 28, 2026
Summary
Federated learning with FedIHRAS enables collaborative AI in radiology, integrating classification, segmentation, and reporting while preserving patient privacy. This approach enhances diagnostic performance and robustness across institutions.
Area of Science:
- Artificial Intelligence
- Medical Imaging
- Federated Learning
Background:
- Federated learning (FL) is a privacy-preserving approach for collaborative AI in healthcare.
- Existing FL frameworks often lack integrated pipelines for complex tasks like classification, segmentation, explainability, and reporting.
- There is a need for comprehensive FL solutions in medical imaging, particularly for multi-institutional radiological analysis.
Purpose of the Study:
- To propose FedIHRAS, a novel privacy-preserving federated learning framework for multi-institutional radiological analysis.
- To integrate multi-task deep learning modules for classification, segmentation, explainability, and automated reporting within a unified FL system.
- To ensure privacy, robustness, and clinical utility in collaborative AI for radiology.
Main Methods:
- Developed FedIHRAS, a federated learning framework integrating multi-task deep learning (ResNet-50, Grad-CAM, SNOMED CT).
- Implemented confidence-weighted aggregation, differential privacy, and secure aggregation for privacy and robustness.
- Evaluated the framework on four large-scale chest X-ray datasets (~874,000 images) across simulated institutional nodes.
Main Results:
- FedIHRAS achieved high diagnostic performance and strong cross-institutional generalization.
- The framework demonstrated improved robustness under non-IID data distributions.
- Experiments confirmed favorable communication efficiency, effective privacy-utility trade-offs, and strong agreement with expert radiologists.
Conclusions:
- Federated learning, exemplified by FedIHRAS, can support scalable, privacy-preserving, and clinically meaningful radiological AI.
- The integrated multi-task learning, explainability, and automated reporting address key limitations of current FL approaches in healthcare.
- FedIHRAS contributes to advancing collaborative AI development in medical imaging and healthcare.
