Towards accelerating multi-energy x-ray ptychographic tomography through deep learning

Abstract

X-ray ptychography is a coherent diffraction imaging technique that enables the visualization of samples with a spatial resolution of up to 4 nm. Spectral capabilities can be integrated into x-ray ptychography to provide chemical speciation of the elements probed within the sample. This spectral information is obtained by performing multiple ptychographic measurements across an energy range around the resonant energy of the probed element. X-ray ptychography can be combined with computed tomography to investigate the internal microstructure of the samples in a non-destructive manner. This is achieved by acquiring two-dimensional (2D) ptychographic projections over an angular range of 0°–180° or 360°. However, ptychographic x-ray computed tomography combined with different incident photon energies may need lengthy acquisition times, sometimes up to a month, depending on the number of 2D ptychographic projections and exposure duration. Reducing the number of projections shortens acquisition time but results in noisier tomograms. In this paper, we demonstrate a method to reduce acquisition time by a factor of four. Our approach involves acquiring only a quarter of the projections required by the Crowther criterion, reconstructing tomograms from this sparse dataset, and subsequently denoising the tomograms using TomoGAN, a generative adversarial network designed for denoising coherent diffraction imaging data. Artifacts introduced by TomoGAN are then corrected by leveraging the redundancy of 2D projections in the spectral dataset. Ultimately, this combination ensures noise suppression and artifact correction, which are the prerequisites for reliable chemical analysis in future studies, within practical acquisition times.