A study of bias and precision in conditional diffusion denoising probabilistic models (DDPMs) for CT imaging

This study investigates bias and precision in conditional Denoising Diffusion Probabilistic Models (DDPMs) for CT imaging. Using the LIDC-LDRI chest CT dataset, we analyze both image- and sinogram-domain models under ideal conditioning. Our findings reveal notable bias and variance, particularly in the sinogram domain, where conventional U-Net architectures struggle to capture long-range features. Even with ideal conditioning, uncertainty may remain too high for certain CT image interpretation tasks. These results underscore the need for improved models and additional physical constraints to mitigate hallucinations and enhance the precision of conditional diffusion models in medical imaging.