Model-driven phase retrieval network for single-shot x-ray Talbot-Lau interferometer imaging.

Abstract

The single-shot x-ray Talbot-Lau interferometer-based differential phase contrast (DPC) imaging is able to accelerate time-consuming data acquisition; however, the extracted phase image suffers from severe image artifacts. Here, we propose to estimate the DPC image via a deep convolutional neural network (CNN) incorporated with the physical imaging model. Instead of training the CNN with thousands of labeled data beforehand, both phantom and biological specimen validation experiments show that high-quality DPC images can be automatically generated from only one single-shot projection image with a certain periodic moiré pattern. This work provides a new, to the best of our knowledge, paradigm for single-shot x-ray DPC imaging.