A physics-informed dual-domain residual neural network for sparse-view CT reconstruction

Although computed tomography (CT) is a highly effective non-invasive diagnostic method,patients are exposed to potential risk from radiation dosage. Therefore, reducing radiation dose while maintaining diagnostic accuracy is a critical issue. One effective approach is decreasing the number of projection views in a CT scan. However, this leads to sparse-view CT problem, resulting in severe artifacts in reconstructed images and posing a significant challenge for diagnosis. In this study, we propose a physics-informed dual-domain reconstruction network (PIDResNet) that integrates processing in both the projection and image domains. Specifically, in the projection domain, we design a physics-based loss function that leverages the internal redundancy of Radon transform unveiled by the recently proposed local correlation equation (LCE). In the image domain, we construct a residual-based convolutional neural network to further improve image quality. Experiments on datasets with different sparsity levels demonstrate the effectiveness of our method from both qualitative and quantitative perspectives.