CNN-based retrieval of x-ray phase-contrast images from experimental data

Speckle-based x-ray phase-contrast imaging (XPCI) offers enhanced sensitivity towards weakly-attenuating materials – such as human tissue – and overcomes limitations of typical phase-contrast imaging systems due to less stringent requirements for setup geometry, coherence, and propagation distances. However, obtaining high-quality phase-contrast images requires accurate tracking of the near-field speckles generated by a random diffuser. A recently proposed speckle tracking method, convolutional-neural-network-based analysis (CADE), offers superior tracking performance, reduced computational times, and robustness to noise when compared to previous state-of-the-art methods but has only been validated on simulated displacements. In this study, we validated CADE on a simulated sine wave sample and experimental XPCI images of a plastic ball. CADE achieved good accuracy in displacement estimation with a simulated sample and further successfully extracted phase contrast signals from experimental imaging data. This brings the implementation of CADE in XPCI one step closer, broadening the potential of XPCI by introducing a novel, improved speckle tracking method.