Single-shot in-line x-ray holography under partially coherent illumination using a self-supervised physics-informed generative reconstruction

Abstract

Single-shot in-line x-ray holography is a promising phase-contrast technique for time-resolved and dose-sensitive synchrotron imaging, but quantitative phase recovery from a single hologram is ill-posed and further complicated by partial coherence. Finite source size, beamline optics, apertures, and propagation geometry reduce Fresnel fringe visibility, and neglecting these effects can lead to reduced phase contrast and systematic reconstruction artifacts. We present a self-supervised, physics-informed reconstruction framework in which the object phase is recovered by optimizing a neural representation to reproduce the measured hologram through a differentiable wave-propagation model. Building on the previously introduced SelfPhish approach, we replace its coherent Fresnel propagator with a differentiable partially coherent forward model. A Gaussian–Schell model is used to demonstrate the approach, although the framework can accommodate more accurate descriptions of synchrotron beam coherence. The measured hologram itself provides the sole supervision, eliminating the need for paired ground-truth data or large training datasets. Simulated and experimental results show that explicitly modeling partial coherence improves quantitative phase reconstruction and suppresses artifacts under realistic beam conditions.