A Semantic Conditional Diffusion Model for Enhanced Personal Privacy Preservation in Medical Images.

Area of Science:

• Medical Image Analysis
• Artificial Intelligence
• Data Privacy

Background:

• Deep learning advances medical imaging but risks patient privacy due to personally identifiable information (PII) in images.
• Protecting PII during medical data transmission is crucial for patient confidentiality.

Purpose of the Study:

• To introduce a novel framework, the Medical Semantic Diffusion Model (MSDM), for synthesizing privacy-preserving medical images.
• To ensure synthetic images maintain the same distribution as original data while removing PII.

Main Methods:

• MSDM utilizes semantic information for image synthesis, encoding it via Adaptive Batch Normalization (AdaBN) into a denoising neural network.
• The Spread Algorithm automatically generates semantic masks, accelerating synthesis and reducing manual effort.
• Experiments were conducted on BraTS 2021, MSD Lung, DSB18, and FIVES datasets.

Main Results:

• MSDM effectively synthesizes medical images, removing PII and ensuring distributional consistency with original data.
• The framework achieved state-of-the-art results across multiple performance metrics on diverse datasets.
• Augmenting datasets with MSDM-generated images improved nnUNet segmentation Dice scores, reaching up to 0.9562.

Conclusions:

• MSDM offers a robust solution for privacy-preserving medical image synthesis.
• The model enhances both image quality and data utility for downstream tasks like segmentation.
• MSDM demonstrates significant potential for secure medical data augmentation in AI research.