ResearchPod Summary
Sparse-view computed tomography (SVCT) is a vital technique for reducing radiation exposure and acquisition time, but the limited number of projection angles makes the reconstruction problem severely ill-posed, leading to significant streak artifacts. While Plug-and-Play (PnP) methods effectively combine data fidelity with learned image priors to mitigate these artifacts, they often require many iterations to converge, limiting their clinical utility. This paper addresses the computational inefficiency of stochastic PnP methods in the context of SVCT.
The authors propose a Multilevel Stochastic Plug-and-Play (ML-SPnP) framework. Multilevel optimization typically accelerates convergence by solving coarse-level problems to compute corrections that are then transferred to the fine-level reconstruction. However, applying this to stochastic PnP is challenging because enforcing prior coherence across resolutions usually requires estimating fine-level prior gradients, which is computationally expensive.
To overcome this, the authors perform the multilevel steps within wavelet multiresolution analysis (MRA) spaces. They demonstrate that, due to the structure of the wavelet decomposition, the prior-coherence correction term vanishes in expectation. This allows the algorithm to bypass the costly estimation of fine-level stochastic prior gradients, retaining only the deterministic data-fidelity correction.
Experiments on medical CT datasets demonstrate that ML-SPnP achieves reconstruction quality comparable to state-of-the-art stochastic PnP methods while substantially reducing the total runtime. The framework successfully integrates stochastic regularization with a multilevel V-cycle, providing a robust and efficient solution for ill-posed inverse problems in medical imaging.
By accelerating the convergence of stochastic PnP methods, this work makes high-quality, iterative CT reconstruction more practical for clinical environments where time and computational resources are constrained. The mathematical insight that prior-coherence corrections vanish in wavelet MRA spaces provides a generalizable strategy for improving the efficiency of other stochastic optimization frameworks in imaging.
AI-generated third-party summary by ResearchPod. Not official content or an endorsement by the paper authors or affiliated organizations.