Structural-preserving diffusion generation for zero-shot CECT-to-NCCT synthesis

Contrast-enhanced and non-contrast computed tomography (CECT/NCCT) are widely used in various clinical scenarios. CECT-to-NCCT synthesis enables NCCT images to be directly generated from CECT images. This streamlines clinical workflows, reduces radiation exposure risk, and prevents co-registration errors. However, to learn the mapping relationship between CECT and NCCT images, existing CECT-to-NCCT synthesis methods typically require supervised settings, which are reliant on substantial paired data. In this paper, we propose a structural preservation diffusion generation (SPDG) method for zero-shot CECT-to-NCCT synthesis. SPDG does not rely on paired data and can be implemented on a pretrained vanilla diffusion model, utilizing an inference-only process to generate NCCT images. Specifically, our SPDG is a guided diffusion generation method that uses the generative prior of the diffusion model and exploits structure-dependent guidance functions to achieve CECT-to-NCCT synthesis. Using guide images from CECT, we design consistency guidance and smoothness constraints as guidance functions for NCCT image generation. Based on the distribution characteristics of the enhanced regions in CECT images, we split the image into a potential enhancement region and an intensity stable region, and designed specific consistency guiding functions for each. The SPDG method is evaluated on abdomen and head and neck datasets, and the results demonstrate that state-of-the-art performance is achieved in unsupervised CECT-to-NCCT synthesis.