X-ray ptychography using physics-enhanced implicit neural representations.

Area of Science:

• X-ray imaging
• Computational imaging
• Materials science

Background:

• Ptychography is a powerful lensless imaging technique.
• Conventional ptychographic reconstruction methods can be sensitive to noise and require ground-truth data.
• Advancements in computational methods are needed for more robust and high-resolution imaging.

Purpose of the Study:

• To introduce a novel deep learning-based approach for ptychographic reconstruction.
• To enhance the stability and image quality of X-ray ptychography reconstructions.
• To enable arbitrary-scale upsampling of reconstructed images.

Main Methods:

• Development of the Deep Ptychographic Iterative Engine (DeePIE) using physics-enhanced implicit neural representations (INRs).
• Representation of the complex object field using two coordinate-based networks for amplitude and phase.
• Joint optimization of networks and probe field directly from diffraction patterns via the X-ray ptychographic forward model.

Main Results:

• DeePIE demonstrated superior reconstruction stability and image quality in both simulations and real X-ray experiments.
• The method successfully reconstructed biological samples using 1.77 nm (700 eV) X-rays.
• Arbitrary-scale upsampling was achieved due to the continuous nature of INRs.

Conclusions:

• DeePIE offers a robust and high-quality solution for X-ray ptychographic reconstruction.
• The method overcomes limitations of conventional techniques, not requiring ground-truth images.
• DeePIE shows significant potential for applications in semiconductor inspection, life sciences, and materials science.