High-resolution interior tomography with a deep neural network trained on a low-resolution dataset

Abstract

Micro-/nano-CT has been widely used in practice to offer noninvasive 3D high-resolution (HR) imaging. However, increased resolution is often at a cost of a reduced field of view. Although data truncation does not corrupt high-contrast structural information in the filtered back-projection (FBP) reconstruction, the quantitative interpretation of image values is seriously compromised due to the induced shifting and cupping artifacts. State-of-the-art deep-learning-based methods promise fast and stable solutions to the interior reconstruction problem compared to analytic and iterative algorithms. Nevertheless, given the huge effort required to obtain HR global scans as the ground truth for network training, deep networks cannot be developed in a typical supervised training mode. To overcome this issue, here we propose to train the network with a low-resolution (LR) dataset generated from LR global scans which are relatively easily obtainable and obtained excellent results.