Yuhe Zhang

Abstract: The development of high-brilliance X-ray sources, such as fourth-generation synchrotron radiation sources and X-ray free-electron lasers, has opened up new opportunities and challenges for X-ray imaging. Addressing the data problem is crucial to fully utilize these facilities. As a data-driven approach, deep learning offers a promising solution to this problem. In this talk, I will present four physics-inspired deep-learning approaches to solve image reconstruction problems in X-ray imaging. First, I will present FFCGAN, a supervised deep-learning approach for shot-to-shot flat-field correction at the European X-Ray Free-Electron Laser (EuXFEL). Second, I will present PhaseGAN, an unsupervised phase reconstruction approach that combines deep learning with the physics of X-ray propagation and interaction with matter, providing a solution for scenarios where conventional phase retrieval approaches fail or are not applicable, such as ultrafast imaging. Last, I will introduce ONIX and 4D-ONIX, two self-supervised approaches to reconstructing 3D and 4D from ultra-sparse (less than eight) projections as measured by X-ray multi-projection imaging (XMPI). XMPI is a rotation-free 3D imaging technique that is compatible with single-pulse approaches but poses a challenge in reconstructing 3D from fewer than ten projections. Thus, the combination of ONIX and 4D-ONIX with XMPI paves the way for 4D X-ray imaging, enabling the recording of 3D movies 1000 times faster than current approaches.