Enhanced sparse view CT reconstruction using ordered subset expectation maximization and asymptotic non-local means algorithms

Abstract This study proposes the Ordered Subset Expectation Maximization–Asymptotic Non-local Means (OSEM-ANLM) algorithm for sparse-view Computed Tomography (CT) reconstruction under limited-data conditions and evaluates its performance across different acquisition settings. The algorithms were evaluated using a digital phantom, a patient chest image, and brain/skull images from the Rando phantom. The acquisition geometry remained a full 180° parallel-beam scan, while the number of uniformly distributed projection views was reduced from 180 to 60, 45, and 30 to simulate sparse-view CT acquisition. The proposed method was compared with the conventional Algebraic Reconstruction Technique (ART). Quantitative assessment employed Peak Signal-to-Noise Ratio (PSNR), Normalized Root Mean Squared Error (NRMSE), and Structural Similarity Index (SSIM), together with visual inspection and linear profile analysis. A SwinUNet deep learning model was developed for comparison with modern AI-based approaches. OSEM-based reconstructions (OSEM reconstruction without post-processing regularization) produced higher-quality images under the evaluated sparse-view conditions than ART, especially at lower view counts. Both OSEM-NLM (OSEM reconstruction followed by NLM regularization) and OSEM-ANLM (OSEM reconstruction followed by ANLM regularization) effectively reduced streak artifacts caused by sparse-view (incomplete projection) acquisition, while OSEM-ANLM achieved superior residual artifact suppression and maintaining detail. For 30 and 60 views, PSNR improved from 30.8710 to 38.8510 to 33.1033 and 40.3210, respectively. The SwinUNet model achieved PSNR, and NRMSE values of 31.6043, and 0.0262 for 30 views, remaining inferior to the proposed method. OSEM outperformed ART, while ANLM reduce sparse-view-induced streak artifacts while maintaining the structural details present in the reconstructed images.