An Alternative Bayesian Reconstruction of Sparse-view CT by Optimizing Deep Learning Parameters

Sparse-view computed tomography (CT) has great potential in reducing radiation dose and accelerating the scan process. Although deep learning (DL) methods have exhibited promising results in mitigating streaking artifacts caused by very few projections, their generalization remains a challenge. In this work, we proposed a DL-driven alternative Bayesian reconstruction method that efficiently integrates data-driven priors and the data consistency constraints. This methodology involves two stages: universal embedding and consistency adaptation respectively. In the embedding stage, we optimize DL parameters to learn and eliminate the general sparse-view artifacts on a large-scale paired dataset. In the subsequent consistency adaptation stage, an alternative Bayesian reconstruction further optimizes the DL parameters according to individual projection data. Our proposed technique is validated within both image-domain and dual-domain DL frameworks leveraging simulated sparse-view (90 views) projections. The results underscore the superior generalization and context structure recovery of our approach compared to networks solely trained via supervised loss.