VVBPConvNet: a lightweight convolutional network for sparse-view CT reconstruction

Sparse-view computed tomography (CT) can effectively reduce radiation dose, but insufficient angular sampling often leads to severe artifacts and loss of fine structural details. Existing reconstruction methods, including conventional filtered backprojection (FBP) and many deep learning approaches built upon it, rely primarily on image-domain representations, where the angular and spatial information contained in individual projections is partially degraded during the backprojection and accumulation process. This information loss limits the quality of sparse-view CT reconstruction. To better preserve and exploit projection information, we propose a View-by-View Backprojection Convolutional Network (VVBPConvNet). Instead of reconstructing images directly from image-domain inputs, VVBPConvNet operates on view-by-view backprojection tensors, enabling the network to access geometric information associated with individual projection views before it is fully merged during reconstruction. To efficiently handle the resulting high-dimensional data, a lightweight network architecture is developed to achieve an effective balance between reconstruction performance and computational cost. Experiments on both simulated and clinical datasets demonstrate that VVBPConvNet consistently outperforms state-of-the-art analytical, iterative, and deep learning methods. The proposed approach achieves effective artifact suppression, improved preservation of fine anatomical structures, and robust performance under sparse-view and noisy conditions. These results suggest that leveraging view-by-view backprojection representations provides a practical and effective strategy for information-preserving sparse-view CT reconstruction.