Federated Learning with Privacy Preserving for Multi- Institutional Three-Dimensional Brain Tumor Segmentation.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computational Biology

Background:

• Accurate brain tumor diagnosis is critical, but data sharing for deep learning is hindered by privacy and legal barriers.
• Federated learning (FL) offers a solution for collaborative model training without direct data sharing.
• Privacy-preserving techniques (PPTs) are essential to ensure data confidentiality in FL.

Purpose of the Study:

• To implement a federated learning approach for brain tumor segmentation.
• To integrate privacy-preserving techniques (PPTs) for enhanced data confidentiality.
• To address challenges in medical imaging data sharing for deep learning models.

Main Methods:

• Utilized a 3D U-Net model trained with federated learning on the BraTS 2020 dataset.
• Incorporated differential privacy as a PPT to safeguard patient data.
• Evaluated segmentation performance using Dice similarity coefficients (DSCs) and 95% Hausdorff distances (HD95) for whole tumor (WT), tumor core (TC), and enhancing tumor core (ET).

Main Results:

• The federated model achieved competitive DSCs and HD95 values on validation and test sets.
• On the test set, DSCs reached 89.85% (WT), 87.55% (TC), and 86.6% (ET), with HD95 values of 22.95 mm, 8.68 mm, and 8.32 mm, respectively.
• Demonstrated the effectiveness of the segmentation approach and its privacy preservation capabilities.

Conclusions:

• A collaborative federated learning model with PPTs successfully segments brain tumor lesions without compromising patient confidentiality.
• The developed model shows high performance in brain tumor segmentation.
• Future work will focus on improving model generalizability and expanding the framework to other medical imaging tasks.