Complex imaging of phase domains by deep neural networks

Area of Science:

• X-ray structural science
• Crystallography
• Computational imaging

Background:

• Phase retrieval is crucial for reconstructing single-particle images from Fourier transform moduli in X-ray structural science.
• Conventional iterative methods for strong-phase objects in Bragg coherent diffraction imaging (BCDI) are slow and initial guess-dependent.

Purpose of the Study:

• To develop a fast and accurate deep-neural-network (DNN) model for single-particle image reconstruction.
• To integrate the DNN model with conventional iterative methods for enhanced accuracy.
• To demonstrate improved convergence using experimental BCDI data.

Main Methods:

• A deep-neural-network model was trained as a universal approximator using synthetic data.
• The DNN model provides a rapid initial estimate of the complex single-particle image.
• The DNN estimate was combined with conventional iterative phase retrieval algorithms for refinement.

Main Results:

• The DNN model offers a fast and accurate estimation of complex single-particle images.
• Combining DNN with iterative methods improves reconstruction accuracy.
• Experimental BCDI data showed enhanced convergence with the hybrid approach.

Conclusions:

• Deep-neural-network models offer a significant advancement in accelerating phase retrieval for X-ray structural applications.
• Hybrid approaches combining DNNs and iterative methods provide a powerful strategy for accurate and efficient image reconstruction.
• This work demonstrates the practical utility of DNNs in advancing Bragg coherent diffraction imaging techniques.