Low-Light Phase Retrieval With Implicit Generative Priors.

Area of Science:

• Scientific Imaging
• Nanotechnology
• Computational Imaging

Background:

• Phase retrieval (PR) is vital for scientific imaging, particularly in nanoscale techniques like coherent diffractive imaging (CDI).
• Low radiation dose imaging is critical for radiation-sensitive samples, but existing PR methods face challenges due to high shot noise in low-dose conditions.
• Current in-situ CDI and data-driven PR methods are not ideal for single-image, data-scarce, or low-dose scenarios.

Purpose of the Study:

• To develop a robust single-image, low-dose phase retrieval method.
• To overcome the limitations of existing phase retrieval techniques in scenarios with high shot noise and limited data.
• To enhance the applicability of phase retrieval for radiation-sensitive samples.

Main Methods:

• Proposed low-dose deep image prior (LoDIP) framework.
• Integration of in-situ coherent diffractive imaging (CDI) principles with implicit generative priors.
• Leveraging deep learning priors for single-image reconstruction in low-dose conditions.

Main Results:

• LoDIP demonstrates superior performance in single-image low-dose phase retrieval compared to existing methods.
• Quantitative evaluations confirm the effectiveness of the proposed approach.
• Successful application of LoDIP in simulated and real experimental settings.

Conclusions:

• LoDIP offers a powerful solution for phase retrieval in low-dose, single-image scientific imaging applications.
• The method effectively addresses the challenge of high shot noise inherent in low-dose imaging.
• LoDIP enhances the utility of techniques like CDI for sensitive samples and data-limited experiments.