Cone-beam CT to synthetic CT translation using conditional 3D latent diffusion-based model

Accurate cone-beam CT (CBCT)-to-synthetic CT (sCT) translation is essential for image-guided adaptive radiotherapy (IGART), where Hounsfield unit (HU) fidelity and structural accuracy directly affect dose calculation. We propose a conditional 3D Latent Diffusion Model (3DLDFM) for volumetric CBCT-to-sCT synthesis. The framework comprises two stages: 1) a 3D variational autoencoder with KL regularization that compresses CBCT volumes into a three-channel latent representation, trained with a composite loss combining L1 reconstruction, perceptual, KL, and adversarial terms; and 2) a conditional 3D U-Net diffusion model that performs iterative denoising in latent space using a DDPM-style noise schedule, conditioned on the input CBCT. We evaluated 3DLDFM on the multi-center SynthRAD2023 dataset comprising 955 paired CBCT/CT volumes spanning head-and-neck, thorax, and abdominal sites. Performance is benchmarked against SwinUNETR, nnUNet, CycleGAN, and Pix2Pix using Mean Absolute Error (MAE), Peak Signal-to-Noise Ratio (PSNR), and Structural Similarity Index Measure (SSIM) within body masks. Across all regions, 3DLDFM achieves the lowest overall MAE (51.40 HU) and the highest overall SSIM (0.9124), while maintaining competitive PSNR (30.60 dB), surpassing all baselines in HU accuracy and structural fidelity. These results demonstrate that the proposed latent diffusion framework provides a robust and generalizable solution for CBCT-to-CT synthesis and strengthens the feasibility of simulation-free adaptive radiotherapy workflows.