MVSS-PCR: A Multiview Subsampling Approach With Prior-Constrained Reconstruction for Self-Supervised CT Denoising

Computed tomography (CT) is widely used for noninvasive diagnostics but exposes patients to radiation. The simplest way to reduce radiation dose is by lowering the tube current and voltage, but this also decreases the number of penetrating photons, increasing noise in the projections. Compared with supervised and unsupervised learning methods, self-supervised learning methods enable noise mitigation. However, existing self-supervised methods often suffer from pixel misalignment during subsampling, leading to excessive smoothing and loss of structural details. To address these challenges, we propose multiview sinogram sampling (MVSS), which enhances alignment between training and target images by aggregating sinograms from multiple views. Furthermore, a nonlocal means (NLM) module with a global search strategy is integrated to overcome the limitations of convolutional neural networks (CNNs) in capturing long-range dependencies, thereby providing higher-quality training targets. Finally, a prior-constrained iterative reconstruction framework with adaptive step-size adjustment is employed to enforce data consistency and restore fine anatomical structures while accelerating convergence. Experimental evaluations on three phantoms and the AAPM dataset demonstrate that the proposed method outperforms competing self-supervised approaches, achieving a PSNR of 35.79 dB and an SSIM of 0.955 on AAPM. Across datasets of Catphan 700, CIRS 057A, and head phantoms acquired on different CBCT systems, the proposed method consistently achieves lower standard deviation (STD) in homogeneous regions with uniform pixel intensity, reaching 28.8 HU, 159.1 HU, and 12.1 HU, respectively, while effectively restoring fine structural details.