Multilevel Stochastic Plug-and-Play for Sparse-View CT Reconstruction

Abstract

Sparse-view computed tomography (SVCT) reduces radiation exposure and acquisition time, but the limited number of projection views makes the reconstruction problem severely ill-posed and leads to streak artifacts when analytical methods are used. Plug-and-Play (PnP) methods combine data fidelity with learned image priors, while stochastic PnP methods use re-noising to reduce the mismatch between denoiser training inputs and reconstruction iterates. However, these methods often require many iterations, which limits their practical efficiency. In this work, we propose multilevel (ML) Stochastic Plug-and-Play (ML-SPnP), a ML stochastic PnP method for SVCT reconstruction formulated on wavelet multiresolution analysis (MRA) approximation spaces. The method combines coarse-to-fine initialization with a finite sequence of ML corrections using resolution-specific denoisers. Under suitable assumptions, the fine-level iterations following the finite ML phase inherit the convergence guarantees of the underlying stochastic PnP algorithm. Experiments on two CT datasets with 20-50 projection views show that ML-SPnP achieves reconstruction quality comparable to the evaluated stochastic PnP methods, while generally requiring less reconstruction time. These results support ML acceleration as an effective strategy for reducing the computational cost of stochastic PnP reconstruction. The code will be made publicly available at https://github.com/antoinedepaepe/ml-spnp.

arXiv comment13 pages, 7 figures, 3 tables