SwinIR-based Dual-Domain Reconstruction for Sparse-View Computed Tomography

Sparse-view computed tomography (CT) remains a significant challenge due to undersampling artifacts and loss of structural detail in the reconstructed images. In this work, we introduce DDSwinIR, a dual-domain reconstruction framework that leverages Swin Transformer-based architectures to recover high-quality CT images from severely undersampled sinograms. DDSwinIR operates in three stages: sinogram upsampling, deep learning-based reconstruction, and a residual refinement module that addresses domain-specific inconsistencies. While previous dual-domain deep learning (DD-DL) approaches improve reconstruction quality, they lack a systematic analysis of component contributions and do not generalize to unseen number of projections. DDSwinIR addresses these gaps through a modular and transparent design, allowing quantification of each network’s module. Our results highlight that early application of data consistency, especially after initial sinogram reconstruction, yields the most substantial and reliable improvements, particularly under extreme sparsity. We also introduce sparse-view concatenation, which enhances performance by improving feature propagation in highly undersampled settings. Extensive evaluation across varying numbers of projections reveal strong generalization when trained on sparse data and tested on denser configurations, but not vice versa, underscoring the importance of low-sparsity training. Compared to conventional reconstruction methods, DDSwinIR achieves superior artifact suppression and detail preservation. This work establishes DDSwinIR as an interpretable and generalizable solution for sparse-view CT, responding to the need for DD-DL reconstruction frameworks for practical applicability.