A tri-contrast signal retrieval neural network for grating-based low-dose X-ray imaging

Grating-based X-Ray imaging (GBXI) can obtain tri-contrast signals of absorption, phase, and dark-field with a laboratory X-ray source and shows a promising future in clinical applications. However, this technique needs a high radiation dose to generate high-quality contrast images. A tri-contrast signal retrieval neural network (TRNN) is presented for low-dose GBXI, consisting of contrast analysis module (CAM) and image enhancement module (IEM). The CAM combines two inception modules of different scales, allowing for more effective extraction of useful information from the phase stepping projections. The IEM is an end-to-end network structure designed to further reduce noise and improve image quality. The TRNN was evaluated using three acquisitions of one mouse obtained at different X-ray tube currents, corresponding to different dose levels. Each group adopted different X-ray tube current, representing different dose level. When the tube current was reduced by 3/4, the feature similarity index measure (FSIM) values of the absorption, phase, and dark-field contrast signals retrieved with the presented TRNN reached 0.9801, 0.9123 and 0.8898 of the standard dose group. When reduced by 7/8, the FSIM values of TRNN reached 0.9703, 0.8876 and 0.8493. By FSIM, TRNN performs best compared with the current signal retrieval techniques. This work has been helpful for the future development and applications of GBXI.