Statistical Penalized-likelihood CT Image Reconstruction with Plug-and-play Priors

Model-based iterative reconstruction (MBIR) methods in CT imaging can provide better image quality and potentially lower radiation dose when compared with analytical methods. Recently, attempts have been made to regularization strategies in MBIR to reconstruct the details while suppressing the noise in the reconstructed images. There are also attempts to denoise CT images with advanced image denoisers though it might introduce artefacts which are difficult to interpret. To combine the strengths of both categories, a recently introduced framework termed Plug-and-play (PnP) is to incorporate advanced image denoisers into the MBIR. Within the PnP framework, in this paper, rather than using least square fidelity term to model the physical aspect of the imaging process, we propose to use statistical penalized-likelihood method where the physical-based Poisson noise model was used. The experimental results confirm the ability of the proposed method in reconstructing the details while retaining the result in an interpretable way.