This PDF file contains the front matter associated with SPIE Proceedings Volume XXXXX, including the Title Page, Copyright information, Table of Contents, and Conference Committee information.

Nondestructive characterization is often carried out on dense materials and objects, such as cargo, and high atomic number advanced manufactured, and traditionally manufactured parts and assemblies. These often require high-energy x-rays, exceeding 1 MeV, and sub-millimeter spatial resolution for characterization. Current high-energy x-ray systems offer limited flexibility in tuning the x-ray energy and spatial resolution. Current x-ray sources include bremsstrahlung tube heads that operate from 10 to 600 kV and accelerators that operate from 2 to 15 MV. The MV systems have spatial resolution on the order of a millimeter which limits the features that can be observed. An alternative approach is to use an inverse Compton scattering (ICS) x-ray source with a linear plasma accelerator (LPA). This is a powerful emerging technology that can provide tunable keV to MeV, quasi-monoenergetic x-rays, and a focal spot size on the order of a micrometer. Our research seeks to answer the question: can ICS x-ray sources fundamentally change the way x-ray non-destructive characterization (NDC) is carried out on these challenging parts, assemblies, and cargo? We performed the first CT scans using the Berkeley Lab Laser Accelerator Hundred TeraWatt (BELLA HTW) ICS x-ray source. We scanned parts of tungsten carbide and steel. Our reconstructed images, although they have artifacts, demonstrate the potential of the LPA-ICS source to image challenging objects. We will discuss the system used, results, lessons learned and paths forward.

X-ray diffraction imaging (XRDI) offers the potential for reduced false alarm rates, increased throughputs, and more sensitive explosives detection performance in aviation security applications. The deployment of computed tomography (CT) systems across carry-on and checked baggage screening lanes has both reinforced the need for orthogonal detection technologies and created an exciting new opportunity for the implementation of XRDI. Our team at Quadridox built a novel XRDI system that, when combined with a CT system, realizes full-tunnel assessment of checked bags at a belt speed of 20 cm/s. We integrated our XRDI system with a Smiths CTX 5800 explosives detection system (EDS) and collected bag data containing both benign and threat objects. We describe the XRDI system, show examples of the resulting hybrid CT and XRD dataset, and present performance results for the hybrid system.

X-ray coherent diffraction-based measurements can provide more specific material information that is complementary to transmission-based material information. With increasing capability of the X-ray coherent diffraction-based systems and recent development of dual modality of X-ray transmission-based and diffraction-based systems, there is a significant potential for improving the overall system threat detection performance for material discrimination. Dual modality systems can yield higher detection probability (Pd) while lowering the probability of false alarm (Pfa), relative to the transmission modality. In this work, we analyze the material discrimination performance for two different machine learning classifiers: support vector machines (SVM) and neural networks (NN), using both simulation and experimental data obtained with a dual-modality X-ray system. Using simulation studies, we demonstrate significant improvement in material discrimination performance afforded by additional complementary information by coherent diffraction for a variety of materials. We further validate these improvements using an experimental dataset collected using real-world objects and materials.

Physics-based synthetic data generation has been shown to complement empirical data in the rapid development of robust algorithms as well as their testing and interpretation. Quadridox’s QSim RT framework can generate rapidly large volumes of realistic X-ray data and associated metadata; while this data can be made to match traditional empirical data collections, the true value of physics-based synthetic data resides in the ways that it goes beyond empirical data to fully explore and map out the parameter space in which the data resides.

We are actively developing methods for augmenting the effectiveness of CT based Explosives Detection Systems (EDS) screening. Our prototype solutions indicate capabilities of improving the security, throughput, and/or TSO deployments at airports of all sizes. Through a collaboration with Sandia National Laboratories, we have developed AI-based explosive detection Automated Threat Recognition Algorithms (ATRs) that can detect both solid and liquid explosives. We leverage the capabilities of deep learning to combine effective atomic number (Z-Eff) and object density, as well as multi-perspective 3D information to precisely localize areas of concerns in luggage. A reduction in false positive detections, and reduced TSA inspection burden is anticipated with the algorithms acting as force multiplier improving security effectiveness. Deep-learning therefore has the potential to revolutionize airport screening. In an airport screening environment, effective implementation of deep-learning will lead to dramatic improvements in efficiency, lower false positive rates, and fewer missed detections.

X-ray coherent scatter yields a weak signal relative to X-ray transmission leading to a challenging reconstruction problem of estimating the scatter spectrum of a material. Here we consider the estimation problem of recovering the spatial mapping of the X-ray coherent scatter spectrum using a spatially modulated (coded-aperture) fan-beam illumination and angle dispersive measurements. We demonstrate that by using both spatial and spectral regularization, within a Bayesian inference framework, we can improve the fidelity of the recovered scatter spectrum from relatively weak scatter measurements. More specifically, in this work we validate this reconstruction performance using experimental measurements obtained from a dual modality X-ray transmission and scatter system. We quantify the reconstruction fidelity our technique relative to maximum-likelihood and group-TV reconstruction techniques.

Despite state-of-the-art deep learning-based computer vision models achieving high accuracy on object recognition tasks, x-ray screening of baggage at checkpoints is largely performed by hand. Part of the challenge in automation of this task is the relatively small amount of available labeled training data. Furthermore, realistic threat objects may have forms or orientations that do not appear in any training data, and radiographs suffer from high amounts of occlusion. Using deep generative models, we explore data augmentation techniques to expand the intra-class variation of threat objects synthetically injected into baggage radiographs using openly available baggage x-ray datasets. We also benchmark the performance of object detection algorithms on raw and augmented data.

Inspection of commerce entering the country is currently a burdensome process, assisted greatly by non-intrusive inspection (NII) technologies such as X-Ray imaging. There are a variety of singular ML techniques that assist in the adjudication process, many are well known, common and widely available with proper data. However, the challenge is to develop a comprehensive system exhibiting certain properties that leverages many techniques simultaneously and combines results to improve overall performance. Here we describe the concepts, components, data and implementation of a system we use for NII adjudication that is well suited to adapting to future needs. We demonstrate several components of this system the benefits of it. Specifically, we show the combination of results from common methods achieves superior performance in concert. First, ensembeling has proven successful in combining results from a variety of object detection models. We show the impact of using our novel graph clique-based method to increase performance over independent models. Additionally, because of performant and reliable object detection results, we demonstrate how constructing an image level contextualization of a scene allows analysis at the metadata level. Lastly, we leverage variational autoencoder (VAE) based similarity models to describe anomalies. We deploy these algorithms in a hierarchical fashion, developing optical signatures for whole NII scans as well as their components (e.g, entire vehicle, trunk, wheels). Here, we quantify the impacts of the inclusion of these methods in image anomaly detection task and demonstrate them in the extensible framework we have developed for NII applications. This paper will serve as an overview of our approach and show results from both combinatorial techniques.

Simulation is a valuable tool for designing and evaluating the performance of x-ray imaging systems. In previous work, a hybrid CT+XRD imaging system was developed for improved identification of threat objects in checked baggage. Through large-scale simulations of this hybrid CT+XRD system, we can investigate the impact of various parameters on system performance. These parameters include varying energy resolution, multi-energy acquisitions, and additional system views. We will report on our findings and evaluate the system performance resulting from these and other variations of the simulated system as well as discuss how these findings may inform future system design.

This study presents an innovative method far from linear density and effective atomic number (Z_eff) based coloring methods in X-Ray Baggage Inspection Systems. Instead of displaying a standard X-Ray image, it is recommended to display several images obtained by parametric modifications in succession called Progressive Sequential Imaging (PSI). The animated images produced by the proposed PSI enable the scanned packages to be examined much shorter than the traditional manual scanning approach through various color or gray-scale image models. Besides, details of the objects, low penetration regions, and materials behind metallic plates are observed more clearly. PSI provides a more powerful and newer imaging method for the X-Ray baggage scanner.

A single-point energy source in X-Ray imaging devices has an artifact of energy loss due to the inverse square rule that also affects the penetration ability. The difference between low and Hi responses is the key feature for effective atomic number estimation in dual-energy X-Ray devices. As the penetration capability alters depending on the location of objects in the devices, the reliability of Z_eff estimates drops. Another reason behind this is the difference in the responses of low and Hi detectors, where Hi read-outs drop faster than Lo read-outs. Thus, the log ratio/Z_eff polynomial modeling estimates lower effective atomic numbers for the same objects as they get further from the generator. This study proposes a novel approach to handle position dependency on atomic number estimation supported by detailed analyses of material type and distance. We model the behavior of different metal/inorganic/organic materials according to their locations in the X-Ray conveyor band and implement a correction mapping approach to preserve the consistency of estimates. Consistency of Z_eff is very important, especially for threat detection, where threat regions are defined as limited Z_eff intervals for dual energy devices. The proposed approach enables material classification consistency against objects’ position on the conveyor band and provides a reliable infrastructure for alarm-based applications in X-Ray devices. The in-depth experiments for various objects, including inert explosives, organic materials, and metallic objects, show the robustness of the proposed band correction approach.

Synthetic data are commonly used to train machine learning models in domains where real data are sparse. In this work, we describe a method to generate synthetic x-ray imaging data by inserting objects into a dual-energy computed tomography scan while simultaneously inserting the beam-hardening and noise artifacts that corrupt real data. This type of data augmentation is useful for training classifiers, for example, by artificially increasing the prevalence of objects of interest in a dataset. This work extends existing 3D Threat Image Projection methods by using dual-energy decomposition to model the energy-dependence of attenuation values in the sinogram data. By summing linear attenuation coefficient functions, objects can be inserted directly into a sinogram while accounting for beam-hardening in the insertion region. In addition, we introduce a calibration method to model the change in noise levels resulting from the insertion of attenuating objects. The performance of the method is demonstrated on a simple phantom scanned with a benchtop microCT system.

X-ray screening can be used to aid in the detection of a variety of threats, including items such as explosives and narcotics. Conventional x-ray imaging provides limited means of distinguishing between different materials. Gratings-based phase contrast x-ray imaging provides images reflecting three distinct materials signatures: absorption, similar to conventional x-ray imaging; phase contrast, proportional to electron density variations; and dark field, which can be related to material texture. These additional signatures may enhance the ability to distinguish materials in x-ray images, leading to detection improvements. Here we discuss considerations for system design and improved detection: approach to measurements at relevant energies for security screening, signature studies of explosives and opioids as well as common benign materials, and development of a prototype screening system.

X-ray Phase Contrast Imaging (XPCI) is an imaging method used to retrieve phase information from an object, thereby revealing more object characteristics, such as refractive index and thickness. Edge Illumination (EI) is a non-interferometric type of XPCI that is spatially dependent. We developed and simulated an alternative type of EI known as Spectrally Responsive Edge Illumination (SREI), which is energy dependent and utilizes an energy-resolving detector. SREI is intended to be more easily implementable than EI due to fewer precision limitations. We will detail results of our SREI experimental investigation and our plan going forward.

Spectral X-ray detectors provide a direct observation of the energy spectrum of a transmitted X-ray beam, both for typical and phase-contrast gratings-based systems. The sensitivity of a gratings-based X-ray system to small-angle deflections from a given particle size is dependent on X-ray energy. Therefore, spectral (energy-sensitive) detectors could be sensitive to particle size, even with a broad spectrum from a commercial X-ray generator. Furthermore, these detectors allow direct observation of how the X-ray spectrum is changing as the beam is passing through an object and gratings, and how this affects grating visibilities used to determine the presence of small-angle deflections. This is a particular issue for higher-energy systems where artifacts from beam hardening are common. We present results exploring the particle-size dependent signatures that are available from a spectral gratings-based phase-contrast X-ray imaging system, and the feasibility of observing them with lab-based, broad-spectrum X-ray generators.

Gratings-based phase contrast imaging provides additional materials signatures that can enhance detection for security screening. However, screening of larger objects requires more penetrating energies, which presents a challenge to gratings-based systems: more penetrating x-rays are more difficult to pattern, while the cross sections for small angle scattering are reduced. We present a study of phase contrast x-ray imaging at endpoint energies above 160 kV, discussing pattern visibility, spectral corrections, and detection of material microstructure. We then discuss considerations for potential implementation.

X-ray phase contrast imaging has the potential to improve image contrast and better differentiate between weakly attenuating materials. Current implementations, however, focus on small biological samples and coherent sources. Here we propose asymmetric illumination as a low complexity variation of x-ray differential phase contrast imaging. With this method, we would utilize angular filtration of the signal at the detector to convert the phase shifts into intensity variation. We will report our findings as we test the feasibility of this method through simulation as well as discuss ongoing efforts in the development of the system.

We have developed a low-cost, scalable detector component technology for x-ray diffraction (XRD) using room-temperature direct-conversion compound semiconductor arrays. Our approach uses two-side, butt-able and stackable modules of cadmium telluride (CdTe) sensors with an edge illuminated geometry. Edge illumination can achieve rapid complete charge collection in a CdTe array when interconnected to custom application specific integrated circuits (ASICs). The modules are tileable in one dimension due to their two-side, butt-able and stackable qualities due to the integration of the ASICs and readout within the active area of the edge illuminated pixels to produce a two-dimensional (2D) field of view (FoV). Since individual panels can be tested before the construction of large 2D FoV detectors, the cost per area is constant. We show results from individual pixels, individual modules, and individual 2D XRD cameras with a FoV up to 6 cm by 6 cm. The cameras demonstrate an energy resolution of ~3 keV full width at half maximum (FWHM) at room temperature across the entire dynamic range from 30 keV to 160 keV with minimal tailing. Edge illumination can also allow a large thickness of the CdTe sensor to be in the incident direction, producing a large detective quantum efficiency (DQE) at high energies (up to 160 keV) without sacrificing charge collection efficiency. Low-cost cameras with large FoVs with sufficiently good energy resolution and DQE could allow XRD imaging to achieve high sensitivity and reduce scan times by reducing false alarms from transmission imaging systems.

We have developed X-ray imaging sensors of the photon-charge counting type that can discriminate X-ray energy, and also developed conventional energy-integrating X-ray imaging sensors. X-ray imaging started in the medical field, but in recent years, it has become necessary to support high-energy X-rays for applications such as nondestructive testing and security inspection. X-ray has a white spectrum, and its maximum energy is determined by the acceleration voltage and the number of photons is determined by the current. When high output is desired for high-throughput imaging, it is common to increase the voltage as well as the current, and good images have been obtained by the skill of the imaging technician. In this study, we discuss the influence of X-ray energy on imaging in terms of the imaging target and the type of X-ray detector, and argue that energy-differentiated imaging with X-rays is important not only for high-performance imaging but also for simple transmission imaging.

This work evaluates two different detector technologies in terms of their performance in making fast, low-signal diffraction measurements. The first detector is a large-area mammography detector that uses a complementary metal-oxide semiconductor (CMOS) crystal, and the second is a cadmium-telluride photon-counting detector. By measuring the diffraction spectra for a diverse range of materials and with acquisition times ranging from 10 seconds and 0.1 seconds, we show how each detector performs as signal-to-noise ratios decrease and counting statistics become less significant. As a result, we show that the photon counting detector slightly better preserves the long-time average signal in short acquisition times in comparison to the CMOS detector when diffraction signals display sharp/narrow features, but that the detectors performed similarly for materials with much broader diffraction signals, like those associated with soft tissue and biological specimen. This leads us to conclude that the photon-counting detector is slightly higher-performing for our purposes.

The unlabeled data are generally assumed to be normal data in anomaly detection (AD) based on semi-supervised learning. This assumption, however, can degrade detection performance when distributions of unlabeled and labeled data are different. To solve the problem, we propose a semi-supervised AD algorithm using KL divergence (SAD-KL) estimating the KL divergence of PDFs of the local outlier factors (LOFs) of the labeled normal and the unlabeled data. We show that the PDFs of the LOFs follow Burr distribution and the SAD-KL shows superior detection probability over the existing algorithms even though it takes less learning time.