Privacy-Preserving Latent Diffusion-Based Synthetic Medical Image Generation.

Area of Science:

• Artificial Intelligence
• Medical Imaging
• Data Privacy

Background:

• Deep learning excels in medical imaging but requires large datasets, posing privacy risks and sharing challenges.
• Diffusion models are advanced AI generative models achieving significant outcomes.

Purpose of the Study:

• To develop a privacy-preserving latent diffusion approach for synthetic medical image generation.
• To enable secure sharing of large medical datasets for deep learning model development.

Main Methods:

• A latent diffusion model was developed to generate synthetic CT, MRI, and PET images.
• The model incorporates a safeguard mechanism for effective patient privacy protection.
• State-of-the-art denoising/super-resolution networks were trained on generated synthetic data.

Main Results:

• Synthetic data generated by the latent diffusion model allowed training of deep learning networks.
• Image quality achieved using synthetic data was comparable to that achieved using original data.
• The embedded safeguard mechanism effectively protected patient privacy.

Conclusions:

• The proposed latent diffusion approach enables privacy-proof sharing of diverse big medical datasets.
• This facilitates the development of deep learning models and potentially federated learning at the data level.