CAFDIM: a group convolution and self-attention fusion-based dual-domain iterative method for sparse-view CT reconstruction

Abstract

Objective. Sparse-view computed tomography (CT) reduces radiation dose and acquisition time by decreasing the number of projection views, but it also makes image reconstruction severely ill-posed, leading to structural distortion and severe artifacts. This study aims to develop an effective reconstruction framework for improving both projection-data fidelity and reconstructed image quality in sparse-view CT. Approach. We propose a group convolution- and self-attention fusion-based dual-domain iterative method (CAFDIM) for sparse-view CT reconstruction. CAFDIM follows a model-informed dual-domain iterative design. The framework consists of the initialization enhancement network, gradient update block, projection-domain repair network, image-domain repair network, and momentum update block. The projection-domain branch employs a deep sparse block to enhance sparse projection features before full-view projection restoration, while the image-domain branch uses edge-guided residual refinement to improve anatomical structure preservation. To enhance local-global feature representation, a Convolution-Attention Fusion Block is embedded into both repair branches by combining group convolution with Pixel Shift Self-Attention. Results. Experiments on simulated and real clinical projection datasets demonstrate that CAFDIM effectively suppresses sparse-view artifacts, preserves anatomical structures, and achieves superior reconstruction accuracy, visual quality, and generalization ability compared with state-of-the-art methods. Significance. CAFDIM provides an effective and efficient dual-domain reconstruction framework for sparse-view CT, showing strong potential for clinical applications in sparse-view and low-dose CT imaging.