Diffusion-Based Synthetic CT Generation from MRI: A Comparative Evaluation with Deep Learning Baselines

Generating synthetic CT images based on magnetic resonance imaging (MRI) is an important research direction in radiotherapy. MRI has good soft tissue contrast but lacks electron density information needed for dose calculation, while CT images can provide Hounsfield Unit information and are therefore often used for radiotherapy planning. This paper compares and analyzes four deep learning methods, Diffusion, U-Net, GAN, and TransformerUNet, for the MRI-to-CT image synthesis task. Experiments are conducted using paired brain MRI-CT data from the SynthRAD2023 Grand Challenge dataset, and the generated results are evaluated using MAE, PSNR, SSIM, and NCC. The results show that the Diffusion model performs best overall, outperforming other models in MAE, PSNR, and NCC. The generated synthetic CT is closer to the real CT and can better preserve anatomical structures. However, the Diffusion model has a slow inference speed and high GPU memory requirements, and further optimization of computational efficiency is needed.