Convolutional analysis operator learning: Application to sparse-view CT : (Invited Paper)

Abstract

Convolutional analysis operator learning (CAOL) methods train an autoencoding convolutional neural network (CNN) in an unsupervised learning manner, to more accurately solve inverse problems. Block Proximal Gradient method using a Majorizer (BPG-M) achieved fast and convergent CAOL, by using sharp majorizers and the momentum terms. This paper proposes a model-based image reconstruction (MBIR) method using autoencoding CNNs trained via CAOL, for sparse-view computational tomography (CT). We apply BPG-M to rapidly and stably solve the corresponding block multi-nonconvex optimization problem. Numerical experiments show that, for sparse-view CT, 1) the proposed MBIR method outperforms the standard MBIR method using edge-preserving regularization; 2) larger parameter dimensions of autoencoding CNNs improve reconstruction accuracy of the proposed MBIR method; and 3) when using BPG-M, sharper majorization is more critical for accelerating its convergence than giving more weights on extrapolation.