Learning to synthesize: robust phase retrieval at low photon counts.

Area of Science:

• Physics
• Computer Science
• Image Processing

Background:

• Deep learning for inverse problems is limited by training data priors.
• Quantitative phase retrieval often suppresses high spatial frequencies, reducing reconstruction quality.
• Existing methods for improving resolution can introduce artifacts and distortions.

Purpose of the Study:

• To develop a new deep learning approach for improved inverse problem solutions, specifically in quantitative phase retrieval.
• To address the suppression of high spatial frequencies in reconstructions.
• To create a method resilient to noise and applicable to various ill-posed inverse problems.

Main Methods:

• A novel "learning to synthesize" (LS) method was developed.
• The LS method learns to process low and high spatial frequency bands separately.
• These frequency bands are then synthesized into full-band reconstructions.

Main Results:

• The LS method produces phase reconstructions with high spatial resolution and without artifacts.
• Reconstructions are robust under high-noise conditions, such as low photon flux.
• The approach effectively overcomes the limitations of frequency band suppression in training data.

Conclusions:

• The LS method offers a significant advancement in quantitative phase retrieval quality.
• This technique is broadly applicable to any inverse problem with uneven frequency band treatment.
• The LS method provides high-fidelity reconstructions in challenging, ill-posed scenarios.