Robust image descriptor for machine learning based data reduction in serial crystallography.

Serial crystallography experiments at synchrotron and X-ray free-electron laser (XFEL) sources are producing crystallographic data sets of ever-increasing volume. While these experiments have large data sets and high-frame-rate detectors (around 3520 frames per second), only a small percentage of the data are useful for downstream analysis. Thus, an efficient and real-time data classification pipeline is essential to differentiate reliably between useful and non-useful images, typically known as 'hit' and 'miss', respectively, and keep only hit images on disk for further analysis such as peak finding and indexing. While feature-point extraction is a key component of modern approaches to image classification, existing approaches require computationally expensive patch preprocessing to handle perspective distortion. This paper proposes a pipeline to categorize the data, consisting of a real-time feature extraction algorithm called modified and parallelized FAST (MP-FAST), an image descriptor and a machine learning classifier. For parallelizing the primary operations of the proposed pipeline, central processing units, graphics processing units and field-programmable gate arrays are implemented and their performances compared. Finally, MP-FAST-based image classification is evaluated using a multi-layer perceptron on various data sets, including both synthetic and experimental data. This approach demonstrates superior performance compared with other feature extractors and classifiers.

Explainable machine learning for diffraction patterns.

Serial crystallography experiments at X-ray free-electron laser facilities produce massive amounts of data but only a fraction of these data are useful for downstream analysis. Thus, it is essential to differentiate between acceptable and unacceptable data, generally known as 'hit' and 'miss', respectively. Image classification methods from artificial intelligence, or more specifically convolutional neural networks (CNNs), classify the data into hit and miss categories in order to achieve data reduction. The quantitative performance established in previous work indicates that CNNs successfully classify serial crystallography data into desired categories [Ke, Brewster, Yu, Ushizima, Yang & Sauter (2018). J. Synchrotron Rad.25, 655-670], but no qualitative evidence on the internal workings of these networks has been provided. For example, there are no visualization methods that highlight the features contributing to a specific prediction while classifying data in serial crystallography experiments. Therefore, existing deep learning methods, including CNNs classifying serial crystallography data, are like a 'black box'. To this end, presented here is a qualitative study to unpack the internal workings of CNNs with the aim of visualizing information in the fundamental blocks of a standard network with serial crystallography data. The region(s) or part(s) of an image that mostly contribute to a hit or miss prediction are visualized.

Dose-efficient scanning Compton X-ray microscopy.

The highest resolution of images of soft matter and biological materials is ultimately limited by modification of the structure, induced by the necessarily high energy of short-wavelength radiation. Imaging the inelastically scattered X-rays at a photon energy of 60 keV (0.02 nm wavelength) offers greater signal per energy transferred to the sample than coherent-scattering techniques such as phase-contrast microscopy and projection holography. We present images of dried, unstained, and unfixed biological objects obtained by scanning Compton X-ray microscopy, at a resolution of about 70 nm. This microscope was realised using novel wedged multilayer Laue lenses that were fabricated to sub-ångström precision, a new wavefront measurement scheme for hard X rays, and efficient pixel-array detectors. The doses required to form these images were as little as 0.02% of the tolerable dose and 0.05% of that needed for phase-contrast imaging at similar resolution using 17 keV photon energy. The images obtained provide a quantitative map of the projected mass density in the sample, as confirmed by imaging a silicon wedge. Based on these results, we find that it should be possible to obtain radiation damage-free images of biological samples at a resolution below 10 nm.

Data reduction for X-ray serial crystallography using machine learning.

Serial crystallography experiments produce massive amounts of experimental data. Yet in spite of these large-scale data sets, only a small percentage of the data are useful for downstream analysis. Thus, it is essential to differentiate reliably between acceptable data (hits) and unacceptable data (misses). To this end, a novel pipeline is proposed to categorize the data, which extracts features from the images, summarizes these features with the 'bag of visual words' method and then classifies the images using machine learning. In addition, a novel study of various feature extractors and machine learning classifiers is presented, with the aim of finding the best feature extractor and machine learning classifier for serial crystallography data. The study reveals that the oriented FAST and rotated BRIEF (ORB) feature extractor with a multilayer perceptron classifier gives the best results. Finally, the ORB feature extractor with multilayer perceptron is evaluated on various data sets including both synthetic and experimental data, demonstrating superior performance compared with other feature extractors and classifiers.

X-ray focusing with efficient high-NA multilayer Laue lenses.

Multilayer Laue lenses are volume diffraction elements for the efficient focusing of X-rays. With a new manufacturing technique that we introduced, it is possible to fabricate lenses of sufficiently high numerical aperture (NA) to achieve focal spot sizes below 10 nm. The alternating layers of the materials that form the lens must span a broad range of thicknesses on the nanometer scale to achieve the necessary range of X-ray deflection angles required to achieve a high NA. This poses a challenge to both the accuracy of the deposition process and the control of the materials properties, which often vary with layer thickness. We introduced a new pair of materials-tungsten carbide and silicon carbide-to prepare layered structures with smooth and sharp interfaces and with no material phase transitions that hampered the manufacture of previous lenses. Using a pair of multilayer Laue lenses (MLLs) fabricated from this system, we achieved a two-dimensional focus of 8.4 × 6.8 nm2 at a photon energy of 16.3 keV with high diffraction efficiency and demonstrated scanning-based imaging of samples with a resolution well below 10 nm. The high NA also allowed projection holographic imaging with strong phase contrast over a large range of magnifications. An error analysis indicates the possibility of achieving 1 nm focusing.

Non-uniform nanosecond gate-delay of hybrid pixel detectors.

A simple experiment to characterize the gating properties of X-ray area detectors using pulsed X-ray sources is presented. For a number of time-resolved experiments the gating uniformity of area detectors is important. Relative gating delays between individual modules and readout chips of PILATUS2 series area X-ray detectors have been observed. For three modules of a PILATUS 300K-W unit the maximum gating offset between the modules is found to be as large as 30 ns. On average, the first photosensor module is found to be triggered 15 ns and 30 ns later than the second and the third modules, respectively.

X-ray deconvolution microscopy.

Recent advances in single-photon-counting detectors are enabling the development of novel approaches to reach micrometer-scale resolution in x-ray imaging. One example of such a technology are the MEDIPIX3RX-based detectors, such as the LAMBDA which can be operated with a small pixel size in combination with real-time on-chip charge-sharing correction. This characteristic results in a close to ideal, box-like point spread function which we made use of in this study. The proposed method is based on raster-scanning the sample with sub-pixel sized steps in front of the detector. Subsequently, a deconvolution algorithm is employed to compensate for blurring introduced by the overlap of pixels with a well defined point spread function during the raster-scanning. The presented approach utilizes standard laboratory x-ray equipment while we report resolutions close to 10 µm. The achieved resolution is shown to follow the relationship [Formula: see text] with the pixel-size p of the detector and the number of raster-scanning steps n.

Connecting structure, dynamics and viscosity in sheared soft colloidal liquids: a medley of anisotropic fluctuations.

Structural distortion and relaxation are central to any liquid flow. Their full understanding requires simultaneous probing of the mechanical as well as structural and dynamical response. We provide the first full dynamical measurement of the transient structure using combined coherent X-ray scattering and rheology on electrostatically interacting colloidal fluids. We find a stress overshoot during the start-up of shear which is due to the strong anisotropic overstretching and compression of nearest-neighbor distances. The rheological response is reflected in uncorrelated entropy-driven intensity fluctuations. While the structural distortion under steady shear is well described by Smoluchowski theory, we find an increase of the particle dynamics beyond the trivial contribution of flow. After the cessation of shear, the full fluid microstructure and dynamics are restored, both on the structural relaxation timescale. We thus find unique structure-dynamics relations in liquid flow, responsible for the macroscopic rheological behavior of the system.

High numerical aperture multilayer Laue lenses.

The ever-increasing brightness of synchrotron radiation sources demands improved X-ray optics to utilise their capability for imaging and probing biological cells, nanodevices, and functional matter on the nanometer scale with chemical sensitivity. Here we demonstrate focusing a hard X-ray beam to an 8 nm focus using a volume zone plate (also referred to as a wedged multilayer Laue lens). This lens was constructed using a new deposition technique that enabled the independent control of the angle and thickness of diffracting layers to microradian and nanometer precision, respectively. This ensured that the Bragg condition is satisfied at each point along the lens, leading to a high numerical aperture that is limited only by its extent. We developed a phase-shifting interferometric method based on ptychography to characterise the lens focus. The precision of the fabrication and characterisation demonstrated here provides the path to efficient X-ray optics for imaging at 1 nm resolution.

Phase unwrapping in spectral X-ray differential phase-contrast imaging with an energy-resolving photon-counting pixel detector.

Grating-based differential phase-contrast imaging has proven to be feasible with conventional X-ray sources. The polychromatic spectrum generally limits the performance of the interferometer but benefit can be gained with an energy-sensitive detector. In the presented work, we employ the energy-discrimination capability to correct for phase-wrapping artefacts. We propose to use the phase shifts, which are measured in distinct energy bins, to estimate the optimal phase shift in the sense of maximum likelihood. We demonstrate that our method is able to correct for phase-wrapping artefacts, to improve the contrast-to-noise ratio and to reduce beam hardening due to the modelled energy dependency. The method is evaluated on experimental data which are measured with a laboratory Talbot-Lau interferometer equipped with a conventional polychromatic X-ray source and an energy-sensitive photon-counting pixel detector. Our work shows, that spectral imaging is an important step to move differential phase-contrast imaging closer to pre-clinical and clinical applications, where phase wrapping is particularly problematic.