Medical imaging deep learning with differential privacy.

Area of Science:

• Artificial Intelligence
• Medical Imaging
• Privacy-Enhancing Technologies

Background:

• Deep learning models require large datasets for medical imaging diagnostics.
• Patient privacy is a critical ethical and legal consideration in medical data handling.
• Differential privacy (DP) offers robust privacy guarantees for sensitive data.

Purpose of the Study:

• To introduce deepee, a novel open-source framework for differentially private deep learning.
• To facilitate the development and deployment of privacy-preserving AI in medical imaging.
• To address the challenges of training deep learning models with sensitive patient data.

Main Methods:

• Developed deepee, a PyTorch-compatible framework implementing differentially private stochastic gradient descent (DP-SGD).
• Utilized parallelized computation of per-sample gradients and efficient memory management.
• Integrated specialized data loading, Gaussian DP-based privacy accounting, and automated DP-conformity adjustments for neural networks.

Main Results:

• Demonstrated successful training of deep learning models with rigorous privacy guarantees on medical imaging datasets.
• Achieved acceptable classification performance for pneumonia detection and excellent segmentation performance for liver tumors.
• Showcased competitive memory consumption and computational performance compared to existing DP frameworks.

Conclusions:

• The deepee framework enables effective and privacy-preserving deep learning for medical imaging tasks.
• It facilitates the wider adoption of privacy-enhancing techniques in healthcare AI research and practice.
• deepee provides a valuable tool for researchers and practitioners navigating the complexities of private AI development.