DROLL: Dual-Domain Reconstruction Network With a High-Fidelity Domain-Transform Operator Based on Learned Low-Rank Prior for Sparse-View CT Reconstruction

Sparse-view CT reconstruction represents a prototypical ill-posed inverse problem. The implementation of deep learning solutions has proven to be highly successful in this field. The dual-domain reconstruction network achieves a favorable trade-off between reconstruction performance and computational cost by leveraging the powerful mapping capability of deep learning and the domain-transform relying on analytical reconstruction algorithms. However, further research is required to enhance the domain-transform in this field. Inspired by the successful utilization of low-rank prior in various medical imaging tasks, we proposed an end-to-end one-shot dual-domain network for sparse-view CT reconstruction. The domain-transform was designed as a high-fidelity multi-channel parallel back-projection in proposed network. In this way, feature maps between channels in the image domain imply strong low-rank priors. We implemented the singular value thresholding algorithm as a network layer, learning parameters and thresholds in a data-driven manner, fully leveraging the low-rank prior across channels to greatly reduce information loss and distortion during domain-transform. Moreover, we constructed a projection completion network based on dual attention mechanism that synthesizes missing view projections by effectively utilizing potential local correlation among projection domains during fan-beam scanning. In the image domain, a refine subnetwork based on Vision Transformer utilizes mix-scale features to implement two-dimensional filtering belonging to the back-projection filter algorithm. Extensive experiments on two clinically relevant datasets have demonstrated that the proposed network achieves competing performance on both quantitative metrics and visual quality.