Ptychographic phase retrieval via a deep-learning-assisted iterative algorithm.

Area of Science:

• Computational imaging
• X-ray microscopy
• Phase retrieval algorithms

Background:

• Ptychography is a powerful computational imaging technique essential for microscopic analysis.
• Phase retrieval algorithms are critical for ptychography's imaging quality.
• Deep neural network (DNN)-based methods offer improved phase retrieval but face limitations with experimental variability and data collection.

Purpose of the Study:

• To develop a robust ptychographic phase-retrieval algorithm overcoming DNN limitations.
• To enhance imaging quality and reduce computational demands in ptychography.
• To improve the adaptability of DNN-based methods to varying experimental conditions.

Main Methods:

• A hybrid approach combining model-based iterative algorithms (e.g., ePIE) with a DNN-based denoiser.
• Training the DNN denoiser using a formula-driven supervised approach with synthetic data, avoiding the need for measured specimen images.
• Evaluating the method using simulations of hard X-ray ptychography and real-world datasets.

Main Results:

• The proposed hybrid method reconstructs higher-spatial-resolution images compared to traditional ePIE and rPIE.
• Achieved comparable or superior image quality with half the number of iterations.
• Demonstrated robustness to hyperparameters and effectiveness with low illumination intensity data.
• Successfully reconstructed images from datasets with lower overlap ratios.

Conclusions:

• The hybrid DNN-iterative algorithm offers a significant advancement in ptychographic phase retrieval.
• This method enhances robustness, reduces data requirements, and improves imaging performance.
• It provides a more adaptable and efficient solution for various ptychographic applications, including those with challenging experimental conditions.