CG-GLORE: New Deep Learning Method for Sparse-View CT Reconstruction

A novel deep learning approach for sparse-view computed tomography (CT) reconstruction has been introduced in a paper on arXiv (ID: 2608.15246). This technique, called CG-GLORE, is a streamlined deep unrolling framework that draws inspiration from second-order optimization. It tackles the challenging inverse problem associated with sparse-view CT, which minimizes radiation exposure by capturing fewer projection views, but can lead to significant streak artifacts. Current deep reconstruction techniques, which depend on first-order updates or extensive regularization networks, may struggle in poorly conditioned scenarios. CG-GLORE employs a conjugate gradient (CG) method to solve a linear system using a structured Hessian surrogate, preserving the physics-based curvature of the data-fidelity term while approximating the learned regularization term with an identity. The framework features a Global-Local Regularization Network to represent image priors. This research, conducted by a team of authors, was presented as a cross-type submission, aiming to enhance reconstruction quality in sparse-view CT and potentially improve medical imaging by allowing lower-dose scans without sacrificing image fidelity.