MR-to-CT Translation Using Frequency-Separated Diffusion Models

Diffusion models have shown great potential in generating high-quality images for magnetic resonance (MR)-to-computed tomography (CT) translation. However, existing diffusion models have limitations in preserving skeletal structures, especially for unpaired datasets. The preservation of skeletal structural details is crucial to avoid disease misidentification and to achieve accurate medical diagnosis and treatment planning. In this paper, a frequency-separated diffusion model (FSDM) is proposed to maintain skeletal structural details, whereby a frequency-separated module can effectively separate the frequency components of medical images in the Fourier domain. In the proposed FSDM, a frequency conversion module is incorporated to convert the MR images into frequency-specific outputs. It leverages the frequency-specific information to guide a subsequent diffusion model for MR-to-CT translation. The results on a public brain dataset demonstrate that our method outperforms the state-of-the-art methods in metrics of Structural Similarity Index Measure (SSIM) and Peak Signal-to-Noise Ratio (PSNR). Our experiments show that FSDM can generate high-quality CT images while preserving skeletal structural details. Our code will be publicly available.