CT Reconstruction using Nonlinear Diffusion Posterior Sampling with Detector Blur Modeling.

There has been a great deal of work seeking to improve image quality in CT reconstruction through deep-learning-based denoising; however, there are many applications where it is spatial resolution that limits application and diagnostics. In this work, we week to improve spatial resolution in CT reconstructions through a combination of deep learning and physical modeling of detector blur. To achieve this goal, we leverage diffusion models as deep image priors to help regularize a joint deblurring and reconstruction problem. Specifically, we adopt Diffusion Posterior Sampling (DPS) as a way to combine a deep prior with a likelihood-based forward model for the measurements. The model we adopt is nonlinear since detector blur is applied after the nonlinear attenuation given by the Beer-Lambert lab. We trained a score estimator for a CT score-based prior, and then apply Bayes rule to combine this prior with a measurement likelihood score for CT reconstruction with detector blur. We demonstrate the approach in simulated data, and compare image outputs with traditional filtered-backprojection (FBP) and model-based iterative reconstruction (MBIR) across a range of exposures. We find a particular advantage of the DPS approach for low exposure data and report on major differences in the errors between DPS and classical reconstruction methods.