Conditional diffusion models for high-fidelity synthetic CT generation from CBCT in nasopharyngeal carcinoma adaptive radiotherapy: a multi-center study.

To develop and validate a conditional diffusion probabilistic model (cDDPM) for generating high-quality synthetic CT (sCT) from cone-beam CT (CBCT) to support accurate online adaptive radiotherapy (ART) for nasopharyngeal carcinoma (NPC). Approach: Planning CT and CBCT images from 1128 NPC patients across two independent centers were retrospectively analyzed. After rigid and deformable registration and preprocessing, 626 cases were used for training, 267 for internal testing, and 235 from another center for external validation. The proposed cDDPM was compared with VAE, GAN, unconditional DDPM, Swin-UNet and Latent Diffusion Model (LDM). Image quality and dosimetric accuracy of sCT were comprehensively evaluated. Main results: In both internal and external validation cohorts, cDDPM-generated sCT significantly outperformed all competing methods across all metrics (MAE, SSIM, MS-SSIM, PSNR, PSNR-hvs, PSNR-hvs-m; all P < 0.001), while showing no significant difference from planning CT (all P > 0.05). Mean dose differences for target volumes and critical organs at risk were < 0.4 % and < 1.6%, respectively, with no statistical significance. 3D gamma passing rates (3 mm/3 %) exceeded 98.0% in both cohorts. Significance: The proposed cDDPM generates sCT images with image quality approaching that of planning CT and introduces negligible dosimetric uncertainty. While prospective clinical validation is still required, the model demonstrates strong potential to facilitate safe and efficient online ART in NPC.