Cone-Beam CT Image Quality Enhancement Using A Latent Diffusion Model Trained with Simulated CBCT Artifacts

A research team led by Naruki Murahashi has published a breakthrough method for dramatically improving the quality of medical Cone-Beam CT (CBCT) scans. CBCT is a faster, lower-dose imaging technique but suffers from low contrast and artifacts (noise) that make diagnosis difficult. The team's novel solution is a conditional latent diffusion model—a sophisticated type of generative AI—specifically trained to clean up these noisy images. The key innovation is its training data: instead of using imperfect real-world CBCT scans, the model learns from perfectly aligned pairs of high-quality CT scans and artificially created 'pseudo-CBCT' scans. These pseudo-scans are made by adding simulated CBCT artifacts to clean CT data, giving the AI a perfect reference for what 'good' looks like.

This self-supervised approach on spatially consistent data allows the model to remove artifacts and enhance contrast without altering the underlying patient anatomy, a critical flaw in previous methods that could 'overcorrect' and change organ shapes. The results are striking. On a test set of 75 cases, the model reduced structural changes (pixel-level anatomical distortions) to less than 1/1000th of those produced by a conventional model trained on real, misaligned image pairs. The processed images also showed a very high correlation (0.916) with the CT value distributions of reference high-quality scans. Furthermore, by operating in a compressed 'latent space', the framework achieves faster processing and superior performance even with constrained training data, making it a practical and efficient tool for clinical deployment.