An End-to-End Deep Network for Reconstructing CT Images Directly from Sparse Sinograms

Recently, deep-learning based methods have been widely used for computed tomography (CT) reconstruction. However, most of these methods need extra steps to convert the sinogrmas into CT images and so their networks are not end-to-end. In this paper, we propose an end-to-end deep network for CT image reconstruction, which directly maps sparse sinogramss to CT images. Our network has three cascaded blocks, where the first block is used to denoise and interpolate the sinograms, the second to map the sinograms to CT images and the last to denoise the CT images. The second block of our network implements the filter backprojection (FBP) algorithm or the Feldkamp-Davis-Kress (FDK) algorithm, where the filter step is implemented by a one-dimensional convolution layer and the backprojection is implemented by a sparse matrix multiplication. By incorporating the FBP/FDK algorithm into our network, training a fully connected layer to convert the sinograms to CT images is avoided and the number of weights of our network is decreased. Our network is trained with two labels, the sinograms and CT images, and can reconstruct good CT images even if the input sinograms are very sparse. Experimental results show that our network outperforms the state-of-the-art approaches on test datasets for the sparse CT reconstruction under fan beam and circular cone beam scanning geometry.