Optimizing Conditional DDPM for Head CT Motion Artifact Reduction: Brain vs. Skull and 3D vs. 2D.

In this study, we introduce a conditional Denoising Diffusion Probabilistic Model (DDPM) approach that employs motion-corrupted images generated by FBP as the condition to reduce motion artifacts in 3D head CT scans. We address two critical questions in this application. First, how can we overcome the disparate performance observed in the skull and brain regions, which is attributable to their distinct intensity ranges? Second, which is more effective for accommodating the 3D nature of head CT and head motion: a 3D or 2D DDPM backbone? The resolution of these questions guides us towards an optimized, image-domain-only DDPM method, demonstrating significant efficacy in reducing motion artifacts in head CT scans.