Privacy-preserving Federated Brain Tumour Segmentation.

Area of Science:

• Medical Artificial Intelligence
• Machine Learning Privacy
• Neuroimaging Analysis

Background:

• Medical data privacy regulations hinder centralized data collection for AI model training.
• Federated learning (FL) addresses this by training models locally and sharing updates, not raw data.
• However, FL model updates can still inadvertently leak sensitive patient information.

Purpose of the Study:

• To investigate the feasibility of applying differential privacy (DP) techniques within a federated learning framework.
• To assess the effectiveness of DP in protecting patient data during federated model training.
• To evaluate the impact of DP on the performance of AI models for brain tumor segmentation.

Main Methods:

• Implementation of federated learning systems for brain tumor segmentation.
• Integration and evaluation of differential privacy mechanisms into the federated learning process.
• Performance assessment using the BraTS dataset, comparing models with and without DP.

Main Results:

• Differential privacy can be applied to federated learning setups for medical data.
• Experimental results demonstrate a clear trade-off between the level of privacy protection and the achieved model performance.
• The study successfully implemented and evaluated practical systems for privacy-preserving federated learning.

Conclusions:

• Differential privacy is a viable approach to enhance patient data protection in federated learning for medical applications.
• Balancing privacy guarantees and model accuracy is crucial when implementing DP in federated learning systems.
• Further research is needed to optimize this trade-off for clinical utility in neuroimaging.