Generation of Synthetic CT Images From MR Images in the Head and Neck Region Using Diffusion Models

In cancer radiotherapy, CT images are essential for planning, while MR images accurately delineate tumors and organs-at-risk, especially in the head and neck (HaN) region. MR-only radiotherapy, with which we generate synthetic CT images from MR data, removes patient radiation exposure. Recent studies indicate that diffusion models yield more realistic images with precise anatomical details and fewer artifacts than generative adversarial networks. In this study, we employ a diffusion model to translate MR images into synthetic CT images for the HaN region. Evaluated on the HaN-Seg dataset of paired CT and MR images of the same patients, our approach achieves a structural similarity index of 92.2%, a peak signal-to-noise ratio of 33.1 dB, and a mean absolute error of 35.3 HU, demonstrating its potential in radiotherapy planning. Validation was extended on a downstream task of organ-at-risk segmentation. Results demonstrate the potential of applying diffusion models into the radiotherapy workflow.