Algorithm for Limited Angle CT Reconstruction with U-net Based Regularization

We propose an algorithm for limited angle CT which extends the LEARN algorithm of Chen et al. [IEEE TMI, 2018]. The algorithm is based on a superiorized [Herman et al., Med Phys, 2012] version of the simultaneous algebraic reconstruction technique (SART), in which a convolutional neural network (CNN) is used to perturb the solution between SART iterations. Specifically, by assuming a fixed number of iterations and unrolling the iteration, it can be trained in end-to-end fashion as a neural network with untrained layers (representing forward and back projection) in addition to trained convolutional layers. Additionally, we replace the three-layer CNN used in LEARN with a U-net-based architecture to achieve multilevel decomposition and precise localization during the iterative training. The algorithm is implemented in Python using the Pyro-NN library [Syben et al., Med Phys, 2019] and trained on simulated limited angle data generated from lung CT images. Our numerical experiments indicate that the superiorized SART algorithm with U-net-based CNN significantly reduces artifacts arising in limited angle scenarios. In particular, it is able to recover internal features that similar methods using the three-layer CNN are unable to recover, such as edges of ribs and soft tissue.