Related Experiment Video
Updated: Jun 6, 2026
05:56
Objectification of Tongue Diagnosis in Traditional Medicine, Data Analysis, and Study Application
Published on: April 14, 2023
Attention U-Net with differential privacy in federated learning framework for brain stroke lesion segmentation
M Adhi Siva1, Chiranji Lal Chowdhary2
1School of Computer Science Engineering and Information Systems, Vellore Institute of Technology, Vellore, Tamil Nadu, 632014, India.
Scientific Reports
|June 4, 2026
Summary
Federated learning with differential privacy (Fed-AttUNet-DP) enhances ischemic stroke lesion segmentation in neuroimaging. This approach ensures data privacy while improving detection accuracy, addressing challenges in fragmented healthcare data.
Area of Science:
- Medical Imaging
- Artificial Intelligence
- Data Privacy
Background:
- Ischemic stroke lesion segmentation is hindered by data privacy concerns and fragmentation across healthcare institutions.
- Centralized approaches risk violating HIPAA and GDPR, while standard federated learning lacks robust privacy and lesion detail.
- Existing methods struggle with non-IID (non-independently and identically distributed) medical data and achieving high segmentation accuracy.
Purpose of the Study:
- To propose Fed-AttUNet-DP, a novel federated learning framework for accurate and privacy-preserving ischemic stroke lesion segmentation.
- To integrate a spatial attention mechanism for improved lesion detection within a federated learning context.
- To ensure dual-layered privacy preservation using client-level differential privacy and server-level secure multi-party computation (MPC).
Main Methods:
- Developed Fed-AttUNet-DP, augmenting federated averaging with attention U-Net and differential privacy (gradient clipping, Gaussian noise).
- Implemented adaptive federated optimization to handle non-IID medical data and accelerate training.
- Conducted experiments on the BRISC2025 dataset involving ten simulated healthcare institutions over 150 communication rounds.
Main Results:
- Achieved high segmentation performance with Dice similarity coefficient (DSC) of 0.930 and Intersection over Union (IoU) of 0.890.
- Demonstrated superior results compared to federated baselines, with minimal accuracy drop versus centralized (0.003) and local-only (0.007) training.
- Reported inference time of 0.8 seconds per volume and a total communication cost of 5.18 GB for 150 rounds.
Conclusions:
- Fed-AttUNet-DP provides formal privacy guarantees, respecting HIPAA/GDPR regulations for medical data.
- The framework effectively mitigates non-IID data heterogeneity across institutions, advancing segmentation accuracy.
- Attention mechanisms, differential privacy, and secure aggregation were validated as key contributors to performance and privacy.