CBCT-to-CT Synthesis: A Hybrid Approach Using CycleGAN and Latent Diffusion

Recent innovations in AI-driven medical imaging have given rise to a powerful hybrid framework that combines CycleGAN and latent diffusion models to synthesize high-fidelity CT images from CBCT data. This methodology addresses long-standing limitations of CBCT imaging, including reduced soft-tissue contrast and spatial resolution, and paves the way for more accurate diagnostic interpretation and treatment planning.

Cone Beam Computed Tomography (CBCT) offers advantages in accessibility and reduced radiation exposure, but its utility is constrained by lower image quality compared to standard CT. Bridging this gap, researchers have explored deep learning models to enhance CBCT imaging. A novel approach published in Nature Communications demonstrated that CycleGAN can effectively translate unpaired CBCT scans into synthetic CT representations, creating anatomically plausible images.

To further elevate realism and detail, latent diffusion models have been integrated into the synthesis pipeline. These models refine the CycleGAN output by learning to reconstruct fine-grained anatomical structures, ensuring clinical features like soft-tissue boundaries and organ interfaces are preserved.

The hybrid synthesis workflow begins with standard preprocessing of CBCT inputs, followed by image translation using CycleGAN—a generative adversarial network suited for unpaired datasets. The preliminary synthetic CT images are then enhanced by latent diffusion processes, which inject high-frequency detail through iterative denoising and latent space reconstruction.

An evaluation published in IEEE Transactions on Medical Imaging showed that this hybrid framework yielded significant improvements in imaging metrics such as SSIM (Structural Similarity Index), PSNR (Peak Signal-to-Noise Ratio), and MAE (Mean Absolute Error). The study highlighted that combining CycleGAN with latent diffusion reduced image artifacts and improved diagnostic confidence.

This two-stage approach has demonstrated consistent performance across datasets. In a recent preclinical study, researchers validated the method against expert radiologist benchmarks and confirmed superior visual and quantitative fidelity compared to single-method synthesis models. Clinically relevant details—such as tumor margins, bony contours, and parenchymal textures—were better preserved in the hybrid outputs.

These results support broader clinical applications, including radiation therapy planning, image-guided surgery, and AI-based diagnostics, where high-resolution imaging is imperative.

The integration of CycleGAN and latent diffusion into CBCT-to-CT conversion offers a promising adjunct to traditional radiologic workflows. By minimizing radiation burden and maximizing anatomical accuracy, this approach could reshape how clinicians interpret CBCT scans in real-time settings.

Nonetheless, wider clinical validation is required. The generalizability of these models across diverse patient populations and imaging protocols remains under investigation. Future studies should also address regulatory considerations for AI-generated medical images and explore embedding such algorithms directly into radiology software suites.

The hybrid CycleGAN–latent diffusion model marks a meaningful step forward in the field of synthetic medical imaging. It significantly enhances CBCT image quality, producing CT-like outputs that retain anatomical accuracy and meet clinical standards. As this technology matures, it holds the potential to redefine non-invasive imaging strategies across radiology and oncology.