Sparse-View CT Reconstruction Based on Dual-Domain Deep Learning

Computed tomography (CT) is an important medical imaging technique widely used in clinical diagnosis. Sparse-view CT is an effective technique for significantly reducing radiation doses in CT imaging, but it often results in severe artifacts when using traditional reconstruction algorithms such as filtered backprojection (FBP). To address this, we propose a dual-domain deep learning architecture for sparse-view CT reconstruction that reduces radiation dose while enhancing image quality. Based on the U-Net model, this method integrates image domain and sinogram domain information through an improved Domain Fusion Module (DFM), which allows early fusion of these features to tackle blurring and artifacts caused by sparse views. Unlike existing methods, we only extract a portion of features for dual-domain fusion while retaining some original features, balancing fusion and information retention. We also employ a Convolutional Block Attention Module (CBAM) in the DFM to prioritize relevant features and improve reconstruction quality. Experiments conducted on the publicly available Mayo2016 dataset demonstrate that our proposed model achieves superior reconstruction quality compared to other state-of-the-art approaches.