Score-based-diffusion-model-based synthetic CT generation from MR images and post-hoc uncertainty analysis

PET/MRI and MR-guided radiotherapy have significantly advanced clinical diagnostics and treatment planning by integrating anatomy and function, while synthetic CT generation from MRI has emerged as a tool for extracting tissue density and attenuation properties. Despite these advances, generating synthetic CT from MRI remains challenging in the pelvic region due to substantial differences in tissue contrast and intensity between the two modalities. Traditional methods using dedicated MR sequences often fail to accurately represent bone structure. Recent deep learning models, including convolutional neural networks (CNNs) and generative adversarial networks (GANs), have improved the quality of synthetic CT generation, they typically lack uncertainty estimation, which limits their clinical utility. In this study, we propose a conditional diffusion model for generating pelvic CT from T1-weigted MR, enhanced by tissue segmentation extracted from TotalSegmentator-MR. We also investigated the self-explanatory characteristics of the diffusion model through uncertainty mapping from multiple realizations without requiring ground truth to enhance interpretability. Our results demonstrate that the diffusion model with tissue segmentation priors, produces higher-quality synthetic CTs and improved tissue differentiation compared to deterministic generative model and diffusion models conditioned only on MR data. Notably, the uncertainty map showed a strong spatial correlation with the ground truth-based error map. Furthermore, the inclusion of segmentation priors also reduces uncertainty, thereby improving the model’s reliability.