DeepPhase: Learning phase contrast signal from dual energy X-ray absorption images

Abstract Due to the high hardware complexity and low dose efficiency of existing X-ray phase contrast imaging, the biomedical and clinical applications of this novel imaging technique have been hindered. This study proposes a deep learning method, named DeepPhase, to extract differential phase contrast (DPC) image from two dual-energy absorption images. It obviates the need of dedicated DPC imaging devices such as Talbot–Lau gratings and is compatible with diagnostic-level dual-energy X-ray imaging hardware. Given two dual-energy absorption images for an object, all we need to produce its DPC image at a certain energy is a well-trained DeepPhase network. Results demonstrate that, compared with conventional Talbot–Lau interferometry, DeepPhase achieves high-quality DPC imaging at multiple dual-energy combinations and low radiation dose.