Self-supervised physics-informed generative networks for phase retrieval from a single X-ray hologram.

Area of Science:

• X-ray imaging
• Phase contrast imaging
• Computational imaging

Background:

• X-ray phase contrast imaging enhances visualization of weakly absorbing structures.
• Propagation-based phase contrast is ideal for dose-critical experiments but requires phase information recovery.
• Conventional phase retrieval methods have limitations in adaptability and require expert parameter tuning.

Purpose of the Study:

• To develop a self-learning approach for phase retrieval using a single intensity measurement (hologram).
• To reconstruct both phase and absorbance of the wave field without relying on training data.
• To overcome limitations of conventional methods in complex or variable experimental conditions.

Main Methods:

• Utilized a physics-informed generative adversarial network (GAN) for phase retrieval.
• Applied the GAN to reconstruct the unpropagated wave field from a single hologram in the Fresnel near-field regime.
• Validated the approach on simulated data and experimental datasets from PETRA III beamline P05.

Main Results:

• The self-learning GAN successfully reconstructed phase and absorbance information from single holograms.
• The method demonstrated robust and consistent performance across diverse imaging conditions and sample types.
• Achieved quantitative, high-quality reconstructions without requiring paired, unpaired, or simulated training data.

Conclusions:

• The proposed physics-informed GAN offers a powerful, data-free solution for X-ray phase retrieval.
• This approach significantly broadens the applicability of phase contrast imaging, especially for in vivo/in situ/operando studies.
• Enables simultaneous retrieval of both phase and absorption information, enhancing quantitative analysis.