High-resolution Interior Tomography with a Deep Neural Network Trained on a Low-resolution Dataset

Micro-/nano-CT has been widely used for noninvasive 3D high-resolution (HR) imaging in many important applications. However, increased resolution is often at a cost of a reduced field of view, which results in many scans being performed with laterally truncated projections. Although data truncation still keeps local details in the projection domain and often allows a high contrast region of interest(ROI) to be reconstructed via filtered backprojection (FBP), the quantitative interpretation of image pixels values is seriously compromised due to the induced shifting and cupping artifacts. State-of-the-art deep-learning-based methods promise accurate and fast 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 dataset generated from low-resolution (LR) global scans, which are relatively easy to obtain. The network learns to predict low-frequency artifacts from LR FBP images reconstructed from truncated LR projections. For inferencing, the LR artifact map is first estimated from down-sampled versions of the HR test image, and then accurately restored to its HR counterpart given the low-frequency nature of the truncation artifacts. In our study, the ROI reconstruction results demonstrate an excellent agreement with the HR ground truth.