Privacy-preserving blockchain-based federated learning for brain tumor segmentation.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Data Security

Background:

• Brain tumor segmentation is crucial for diagnosis and treatment but hindered by data privacy concerns and model trust issues.
• Current data sharing methods are limited by the sensitive nature of health information, preventing collaborative model training.
• Accurate segmentation is challenging due to tumor heterogeneity, irregular shapes, and variable locations.

Purpose of the Study:

• To propose a novel framework addressing privacy and trust challenges in medical data sharing for AI model training.
• To enable collaborative training of a global brain tumor segmentation model without compromising patient privacy.
• To enhance the accuracy and reliability of brain tumor segmentation through secure, decentralized learning.

Main Methods:

• Implementation of a federated learning approach for decentralized model training across multiple institutions.
• Integration of a permissioned blockchain to securely share encrypted model parameters and aggregate gradients.
• Development of a masking technique to preserve the privacy of individual model parameters during sharing.

Main Results:

• The proposed blockchain-federated learning framework demonstrated significant improvements in brain tumor segmentation.
• Achieved a 1.99% increase in Dice similarity coefficient for enhancing tumors.
• Reduced Hausdorff distance for whole tumors by 19.08% compared to baseline methods.

Conclusions:

• The framework effectively addresses privacy and trust concerns in medical data sharing for AI.
• The approach enhances the performance and reliability of brain tumor segmentation models.
• Enables collaborative learning, benefiting all participating healthcare entities with improved diagnostic capabilities.