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

Deep learning has significantly advanced medical image processing, yet the inherent inclusion of personally identifiable information (PII) within medical images—such as facial features, distinctive anatomical structures, rare lesions, or specific textural patterns—poses a critical risk to patient privacy during data transmission. To mitigate this risk, we introduce the Medical Semantic Diffusion Model (MSDM), a novel framework designed to synthesize medical images guided by semantic information, synthesis images with the same distribution as the original data, which effectively removes the PPI of the original data to ensure robust privacy protection. Unlike conventional techniques that combine semantic and noisy images for denoising, MSDM integrates Adaptive Batch Normalization (AdaBN) to encode semantic information into high-dimensional latent space, embedding it directly within the denoising neural network. This approach enhances image quality and semantic accuracy while ensuring that the synthetic and original images belong to the same distribution. In addition, to further accelerate synthesis and reduce dependency on manually crafted semantic masks, we propose the Spread Algorithm, which automatically generates these masks. Extensive experiments conducted on the BraTS 2021, MSD Lung, DSB18, and FIVES datasets confirm the efficacy of MSDM, yielding state-of-the-art results across several performance metrics. Augmenting datasets with MSDM-generated images in nnUNet segmentation experiments led to Dice scores of 0.6243, 0.9531, 0.9406, and 0.9562 underscoring its potential for enhancing both image quality and privacy-preserving data augmentation.