Invertible Learned Primal-Dual

Jevgenija Rudzusika, Buda Bajic, Ozan Öktem, Carola-Bibiane Schönlieb, Christian Etmann

Keywords: iterative reconstruction, invertible neural networks, deep learning, inverse problems, computed tomography

TL;DR: We propose an invertible Learned Primal-Dual architecture for 3D reconstruction in computed tomography.

Abstract: We propose invertible Learned Primal-Dual as a method for tomographic image reconstruction. This is a learned iterative method based on the Learned Primal-Dual neural network architecture, which incorporates ideas from invertible neural networks. The invertibility significantly reduces the GPU memory footprint of the Learned Primal-Dual architecture, thus making it applicable to 3D tomographic reconstruction as demonstrated in the experiments.