Computation Overhead Optimization Dual-Domain Network with Multi-Level Cross-Domain Connections for Sparse-view CT Reconstruction

To reduce the harmful effects of radiation dose on the human body, the sparse scanning method is widely used. Recently, deep learning methods have been widely used in rather challenging reconstruction tasks such as sparse angles, and the dual domain method has a natural advantage because it can process the information in both the sine domain and the image domain. However, the existing two-domain methods do not pay enough attention to the allocation of training costs and strategies for the two domains. To solve this problem, in this paper, we design a Computation Overhead Optimization Dual-Domain Network with Multi-Level Cross-Domain Connections for Sparse View CT Reconstruction. We carries out feature extraction and image recovery for projection and image respectively through two Transformer-based subnets of coding-decoding type. In particular, in the two domains, we use the multilevel cross-domain connection method to connect the coding layers of the same scale in the two subnetworks, and adopt the two-channel method for up-sampling .The experimental results on clinical data validate the effectiveness of our method and training strategy. From the subjective and objective evaluation indexes of reconstruction results, our method can outperform some of the most advanced methods in the past.