This PDF file contains the front matter associated with SPIE Proceedings Volume 10999, including the Title Page, Copyright information, Table of Contents, Author and Conference Committee lists.

Since 2013, a Duke University / University of Arizona collaboration has been investigating how to optimize the hardware design of aviation-security x-ray systems. The goal of the effort is to develop a detection-algorithmagnostic approach that focuses on the ability of the hardware to capture threat/non-threat information in the transduced measurements. The resulting approach combines high-fidelity, high-throughput simulation of large numbers of synthetic bags with information-theoretic based metrics and allows trade studies that vary key system hardware parameters (e.g. spectral resolution, number of views, etc.). In the intervening years, this framework has been continually expanded and side collaborations with various OEMs have been initiated to use the tool to explore areas of design space relevant to their interests. In this talk I will discuss the history of this effort as well as the current status, and will speculate on where this approach can go in the future.

We present Spectral X-ray Computed Tomography (SCT) estimations of material properties directly from energy-dependent measurements of linear attenuation coefficients (LAC). X-ray Computed Tomography (CT) is commonly utilized to characterize the internal properties of an object of interest. Dual-Energy X-ray CT allows material characterization into energy-independent physical properties such as Z_e and electron density ρ_e. However, it is not robust in presence of dense materials and metal artifacts. We report on the performance of a method for system-independent characterization of materials that introduces a spectroscopic detector into X-ray CT, called spectral ρ_e/Z_e estimation (SRZE). We benchmark the SRZE method against energy-integrated measurements in material classification tests, finding superior accuracy in the predictions. The advantage of this technique, over other methods for material characterization using x-ray CT, is that it does not require a set of reference materials for calibration. Moreover, the simultaneous detection of spectral features makes it robust to highly attenuating materials, since the energy intervals for which the attenuation is photon limited can easily be detected and excluded from the feature estimation.

Simulations of x-ray scanners have the potential to aid in the design and understanding of system performance. We have previously shown the usefulness of a high-throughput simulation framework in pursuit of information theoretic analysis of x-ray systems employed for aviation security. While conclusions drawn from these studies were able to inform design decisions, they were limited to generic system geometries and na¨ıve interpretations of detector responses. In collaboration with the SureScan Corporation, we have since expanded our analysis efforts to include their real world system geometry and detector response. To this extent, we present our work to simulate the SureScan x1000 scanner, a fixed-gantry spectral CT system for checked baggage. Our simulations are validated in terms of system geometry and spectral response. We show how high fidelity simulations are used with SureScan reconstruction software to analyze virtual baggage. The close match between simulated and real world measurements means that simulation can be a powerful tool in system development. Moreover, the close match allows simulation to be a straightforward avenue for producing large labeled datasets needed in machine learning approaches to automatic threat recognition (ATR).

Established non-destructive, but penetrating techniques for baggage inspection are primarily based on the transmitted X-ray signal (AT systems, Computed Tomography), which allows determining the objects’ density and average atomic composition. X-ray diffraction is one of the few options to combine the penetrating properties of X-rays with the opportunity to gather material-specific information about the investigated object. While X-ray diffraction for scientific applications is widely used to determine spatial information on molecular level, the application in real-time at high throughput –as required for baggage inspection- has been emerging technology for a few decades now. This presentation will give an overview of the current established and evolving technologies and the required key performance parameters for the application in baggage screening. One realization of a fully 3D baggage screening concept named “X-ray Diffraction Imaging” (XDi) will be described in more detail.

Transmission-based imaging and X-ray diffraction-based material analysis have largely developed independently. However, for a variety of applications ranging from in-vivo soft tissue analysis to concealed explosives detection, it is necessary to realize high-fidelity, spatially-resolved material discrimination. We therefore seek to understand to what degree transmission and X-ray diffraction (XRD) complement one another and can be implemented practically, particularly in the case of explosives detection in aviation security. Using a combination of simulated and experimental data, we identify the relative value of the X-ray signatures available to transmission and XRD measurements, and explore how the measurement fidelity can impact these results.

X-ray diffraction tomography (XRDT) is an exciting imaging modality because of its capacity to combine volumetric imaging with material discrimination. Despite this fact, practical implementation of such systems to security workflows have been fraught with challenges due to real world" constraints” one of the most notable being a limitation on the total photon budget. This challenge is exacerbated given that one of the more pronounced trade-offs in XRDT system design is that between imaging performance and the quantity of relevant, scattered photons collected. Consequently, two approaches have emerged that operate at different ends of this trade-off continuum. At one extreme, direct tomography (DT) uses a high degree of collimation to realize robust resolution but suffers from low signal levels. At the other extreme, coded aperture XRDT (CA-XRDT) employs coded apertures which allow for detection of a substantially greater number of photons but at the cost of signal multiplexing and potentially less robustness to object extent. While these two systems differ only slightly in hardware (i.e. whether a collimator or coded aperture is used), a fair and fundamental comparison between the two is not straightforward and has never been performed. Furthermore, such an analysis is important for understanding the strengths and weaknesses of each system and thereby identifying an optimal architecture. In this paper, we first present our methodology to define the theoretical resolution of DT and CA-XRDT systems, focusing on the case of a pencil beam geometry. After using this approach to tweak system design in simulation such that both systems have the same theoretical average resolution, we then conduct a numerical investigation of the imaging performance of each system as a function of the measurement SNR and target object width for linear and area array detector geometries. We find in our particular simulation study that, while there are cases where both systems can identify and localize an object equally well, there are certain imaging scenarios where CA-XRDT outperforms DT and vice versa. In addition, we find that DT generally provides more information per photon than CA-XRDT but that there can be comparatively less information overall.

We report on the performance improvement in Automatic Threat Recognition (ATR) algorithm through the incorporation of self-tuning spectral clustering and a convolutional neural network texture model (CNN). The self-tuning clustering algorithm shows the ability to vastly reduce the amount of bleedout in threat objects resulting in better segmentation and classification. The CNN texture model shows improved detection and classification of textured threats. These additions have markedly improved the ATR. The tests performed using actual CT data of passenger bags show excellent performance characteristics.

Transmission x-ray systems rely on the measured photon attenuation coefficients for material imaging and classification. While this approach provides high quality imaging capabilities and satisfactory object discrimination in most situations, it lacks material-specific information. For airport security, this can be a significant issue as false alarms require additional time to be resolved by human operators, which impacts bag throughput and airport operations. Orthogonal techniques such as X-ray Diffraction Tomography (XRDT) using a coded aperture provide complementary chemical/molecular signatures that can be used to identify a target material. The combination of noisy signals, variability in the XRD form factors for the same material, and the lack of a comprehensive material library limits the classification performance of the correlation based methods. Using simulated data to train a 1D Convolution Neural Network (CNN), we found relative improvements in classification accuracy compared to the correlation based approach we used previously. These improvement gains were cross-validated using the simulated data, and provided satisfactory detection results against real experimental data collected on a laboratory prototype.

Task-specific adaptive sensing in computed tomography (CT) scan is critical to dose reduction and scanning acceleration. Due to the sequential nature of the CT acquisition process, the information of the objects aggregates as the measurement process progresses. Conventional adaptive sensing methods, aiming to maximize the task-specific information acquisition, formulate the measurement strategy as an optimization problem with assumptions in object distributions (for example, Gaussian mixture model), which requires considerable computational time and resource during the acquisition. In our work, we propose a machine learning approach to learn task-specific data-acquisition policy, with the only assumption on the locality and composition of the objects, which shifts the computation load to the pre-acquisition stage. We analyze our learned method on public dataset comparing to a stochastic policy which plans the acquisition randomly and a uniform policy which plans the acquisition with a fixed interval. Based on our experiments the learned method requires at least 25% fewer acquisition steps than the stochastic and uniform policies.

The material-specific information contained in X-ray diffraction (XRD) measurements make it attractive for the detection of threats in airport baggage. Spatially-localized XRD signatures at each voxel in a bag may be obtained with a snapshot via coded aperture XRD tomography, but measurement unceratinty due to data processing and low SNR can lead to loss in information. We use machine learning and non-linear dimension reduction to identify threat and non-threat items in a way that overcomes these variations in the data. We observe the emergence of clusters from the data, possibly providing new prospects for XRD-based classification. We further show improved performance using machine learning methods relative to a conventional, correlation-based classifier in the low-SNR regime.

X-ray diffraction (XRD) tomography continuous to be a promising technology for maintaining a high detection probability with low false alarm rates while adding new threat classes. By collecting the diffracted (coherently scattered) X-rays, one realizes several key advantages over transmission-based systems: • Measure several additional features by which to identify material composition • Tomographic 3D spatial imaging with only a single view • Automatic, PC based, explosives detection algorithms for replacing human operators A key requirement for XRD technology is excellent energy resolution (ER) of the detectors used in the scanner. Existing spectroscopic detectors offer low enough ER (below 6keV) but, unfortunately, operate at rather low count rate (typically under 1kcps/mm2). As a result, commercial XRD scanners, such as the XRD3500, require long scan times. As a result, it is difficult to effectively use these scanners in airports. A breakthrough in XRD technology was achieved through the use of coded apertures, which increase the signal amplitude by 2-3 orders of magnitude compared to traditional heavy collimated systems. While the brighter resulting XRD signal requires complex signal processing to eliminate increased scatter, it has been shown to produce a much faster XRD scanner response (seconds instead of minutes). A practical implementation of this new approach requires high count rate (between 1kcps/mm2 and 1Mcps/mm2) while maintaining very low ER (below 6keV) and sub-mm spatial resolution required for angular detection precision. In the last 5 years, Redlen Technologies has developed high-flux CZT detection technology for medical Computed Tomography (CT) that is currently being deployed by major medical OEMs into clinical applications. The technology is based on a 22x34 [748 pixels] 2-D array with a pixel pitch of 330um. The associated high-speed photon counting ASIC that allows for event detection operates up to 250Mcps/mm2. Recently we have found a way to reconfigure that detector technology platform into an XRD platform. In this paper we will present experimental results of our 2-D 22x32 CZT pixel array that is currently available for deployment into XRD scanner platforms. The CZT sensors used in this platform are 2mm thick with a 330 um pixel pitch and operate without polarization up to 250Mcps/mm2. In the CT mode, the detectors operate in the 16-190 keV range with energy resolution of 6.9 keV and standard deviation of 0.7 keV across 748 pixels. In the XRD mode, the detectors operate in the 12-150 keV range with energy resolution of mean value of 5.6keV and standard deviation of 0.6keV across 748 pixels. We believe these performance levels are more than sufficient to enable operating XRD scanner at the optimum performance levels.

Given the daunting task of developing novel/improved methodologies for explosives detection equipment, the government has taken a novel approach. The use of industry and university collaboration has aided in propelling technology. The ongoing interaction has proven to move toward viable solutions sooner than expected. Given this pace of development, the question remains, how do we evaluate, certify, deploy, and maintain them operationally?

Detecting material anomalies in baggage requires a high-throughput X-ray measurement system that can reliably inform the user/classifier of pertinent material characteristics. We have developed a comprehensive high-fidelity simulation framework capable of modeling a multi-energy X-ray fixed gantry computed tomography transmission system. Our end-to-end simulation framework includes experimentally validated models of sources and detectors, as well as virtual bags to emulate the X-ray measurements generated by the fixed gantry X-ray CT system. This simulation capability enables us to conduct exploratory system trade-off studies around the current fixed gantry system, in terms of the source detector geometry, detector energy resolution and other relevant system parameters to assess their impact on the threat detection performance. Using scalable information-theoretic metrics, evaluated on simulated system data, we are able to provide quantitative performance bounds on the performance of the candidate system designs. In this work, we will report results of our initial system design trade-off studies focused on detector energy resolution and energy partitioning and how they impact the threat detection performance.

Extending our prior work, we propose a multi-energy X-ray measurement model incorporating material variability with energy correlations to enable the analysis and exploration of the performance of X-ray imaging and sensing systems. Based on this measurement model we provide analytical expressions for the Cauchy-Schwarz mutual information (I_CS) measure that quantifies the performance limits of an X-ray measurement system for the threat-detection task. We analyze the performance of a prototypical X-ray measurement system to demonstrate the utility of our proposed material variability measurement model.

X-ray Computed Tomography (CT) based systems are widely used for security and industrial inspection applications like passenger baggage and cargo screening at the airports, shipping container screening at ports, and industrial parts inspection. Typically, such CT systems employ a mechanical rotating gantry, with a X-ray source and detector array(s), to collect a large number of angularly diverse projections scans of an object. However, such a mechanical scanning mechanism adds to the system cost and maintenance. In this work, we consider the next generation Rectangular-Fixed-Gantry (RFG) CT system architecture containing several X-ray sources and detector arrays deployed in a fixed geometry. As such, such a RFG CT system architecture affords us the opportunity to explore non-standard multiplex measurement designs employing simultaneous illumination from multiple sources. The goal of our multiplexed measurement design is to minimize the bag reconstruction error (or Mean Square Error (MSE)) for a given source flux budget and/or fixed measurement time. We utilize a Bayesian Cramer-Rao Bound (CRB) on reconstruction error as a multiplex measurement design metric, subject to a fixed source flux/measurement time constraint. We find that the resulting optimized multiplexed measurement designs can significantly outperform the sequential measurement design employed in a traditional RFG CT system. We quantify this reconstruction fidelity improvement with multiplexed measurement design using simulation studies and assess its potential benefits in terms of operational system performance metrics such as bag reconstruction fidelity and system throughput.

For understanding the interaction between projectiles and target structures at impact velocities up to 2000 m/s and higher, a fast-multi-flash X-ray system is needed to look through materials. For that reason a fast optical diagnostic system was developed which is equipped with a multi-flash X-ray system, a fast decay scintillator screen and a fast gating intensified multi-frame CCD camera. The already low parallax caused by the compact annular installed RXanodes was eliminated by mathematical affine transformation. A first basic version of an image processing and analysing software was developed to eliminate the already low parallax, to improve the image quality, detect fragments and debris, as well as calculating the object velocity. This high-speed diagnostic system was used for ballistic testing to investigate the ballistic performance of composite body armour, ballistic helmets and laminated armour.

Prior to the advent of modern deep learning techniques, data mining was already being used for image processing in aviation security. In 2010, the paper “Applying data mining to false alarm reduction in an aviation explosives detection system”, detailed lessons learned from using automated data mining techniques for false alarm identification. The paper included a series of observations and recommendations. Nearly a decade later, deep learning is showing tremendous promise for a variety of image processing problems (in general) and to CT-based explosives detection systems (EDS) in particular. While some risks and shortcomings of deep learning are understood, the particular issues associated with aviation security applications may not be. We revisit the earlier work and see whether it withstands the test of time and still applies. We then combine the earlier work with modern deep learning design guidelines, to form a guide to using deep learning for aviation security.

X-ray computed tomography has been recently applied to capture the dynamic behaviors of complex material systems in 4D. The dynamic 3D acquisition, however, usually leads to insufficient data acquisition with low-dose X-ray radiation and limited-angle projections. A high-fidelity CT reconstruction is challenging based on the severely limited acquisition. While prior constraint, such as local smoothness, can improve the quality of reconstructions, a more general reconstruction strategy to include structural features on a range of different scales proves to yield better reconstruction results and are more adaptive to complex structured materials. In this work, we develop the hierarchical synthesis network to establish structural priors for sparse-view CT reconstruction, which achieves high-fidelity with an improved computation efficiency. We found that the established knowledge of structural priors on each different scale can be independently transferred to sparse-view CT reconstruction under different conditions, enabling the transfer of non-local features into the reconstruction of a phase tomography application.

The paper proposes an algorithm to improve the accuracy of searching the boundaries of objects. We analyze X-ray images. Most often, images obtained in the X-ray range are subject to distortion. Noise and blurring are caused by the nature of the source and the imperfection of the sensor. In our work, we propose the use of a multicomponent processing algorithm. This type of processing is based on step-by-step image analysis. At the first stage, the stage of image blur recovery is performed (deblurring). At the next stage, the operation of filtering images (denoising) and processing the boundaries of objects is performed. The first two steps are performed using adaptive local processing with nonoverlapping windows. At the next stage, the multicriteria processing method is used for the one-dimensional and two-dimensional signals. The first approach is used to reduce the effect of the noise component at the boundaries of objects and is also used as a boundary detector. The second criterion is used to reduce the effect of the noise component in locally stationary areas. The efficiency of the proposed algorithm is shown using the example of medical X-ray data processing and the results of computed tomography. Using the example of the developed software for the analysis of CT images and the restoration of the lost elements of the bone structure, an example of the application of the proposed approaches for performing primary data processing operations is shown.

We present a method for the automated detection of firearms in cargo x-ray images using RetinaNet. RetinaNet is a recently proposed powerful object detection framework that is shown to surpass the detection performance of state-of-art two-stage R-CNN family object detectors while matching the speed of one-stage object detection algorithms. We trained our models from scratch by generating training data with threat image projection (TIP) that alleviates the class imbalance problem inherent to the x-ray security inspection and eliminates the need for costly and tedious staged data collection. The method is tested on unseen weapons that are also injected into unseen cargo images using TIP. Variations in cargo content and background clutter is considered in training and testing datasets. We demonstrated RetinaNet-based firearm detection model matches the detection accuracies of traditional sliding-windows convolutional neural net firearm detectors while offering more precise object localization, and significantly faster detection speed.

The objects in the medical images are not visible due to low contrast and the noise. In general, X-ray, computed tomography (CT), and magnetic resonance imaging (MRI) images are often affected by blurriness, lack of contrast, which are very important for the accuracy of medical diagnosis. It is difficult to segmentation in such case without losing the details of the objects. The goal of image enhancement is to improve certain details of an image and to improve its visual quality. So, image enhancement technology is one of the key procedures in image segmentation for medical imaging. This article presents a two-stage approach, combining novel and traditional algorithms, for the enhancement and segmentation of images of bones obtained from CT. The first stage is a new combined local and global transform domain-based image enhancement algorithm. The basic idea of using local alfa-rooting method is to apply it to different disjoint blocks of different sizes. We used image enhancement non-reference quality measure for optimization alfa-rooting parameters. The second stage applies the modified active contour method based on an anisotropic gradient. The simulation results of the proposed algorithm are compared with other state-of-the-art segmentation methods, and its superiority in the presence of noise and blurred edges on the database of CT images is illustrated.

Gratings-based x-ray imaging can provide additional materials signatures, including refraction which is proportional to variations in electron density, and scatter which is sensitive to sub-resolution texture. Phase contrast measurements have been conducted using a variety of approaches, including Talbot-Lau interferometry, coded aperture systems, and single absorption grid systems. Because of the simultaneous requirements for fine spatial patterns to detect small angular changes, and the thickness of material required to modulate a penetrating beam, many phase contrast measurements are conducted at relatively low energy, below 100 kV. Many applications in security screening require higher energies in order to penetrate larger objects. Here, we use a single absorption grid with direct imaging of the projected pattern to perform phase contrast measurements. A second grid is used for a beam hardening correction. We present measurements of pattern visibility as a function of energy up to 450 kV, demonstrating that the necessary beam patterning can be extended to higher energies. We also present measurements of a textured and homogeneous material as a function of energy, demonstrating that a texture signature is still present as energy is increased, and that the beam-hardening correction correctly accounts for and removes spectral effects on pattern visibility. To the best of our knowledge, this represents the highest energy demonstration of this technique to date, and enables new application areas.

High-quality image products in an X-Ray Phase Contrast Imaging (XPCI) system can be produced with proper system hardware and data acquisition. However, it may be possible to further increase the quality of the image products by addressing subtleties and imperfections in both hardware and the data acquisition process. Noting that addressing these issues entirely in hardware and data acquisition may not be practical, a more prudent approach is to determine the balance of how the apparatus may reasonably be improved and what can be accomplished with image post-processing techniques. Given a proper signal model for XPCI data, image processing techniques can be developed to compensate for many of the image quality degradations associated with higher-order hardware and data acquisition imperfections. However, processing techniques also have limitations and cannot entirely compensate for sub-par hardware or inaccurate data acquisition practices. Understanding system and image processing technique limitations enables balancing between hardware, data acquisition, and image post-processing. In this paper, we present some of the higher-order image degradation effects we have found associated with subtle imperfections in both hardware and data acquisition. We also discuss and demonstrate how a combination of hardware, data acquisition processes, and image processing techniques can increase the quality of XPCI image products. Finally, we assess the requirements for high-quality XPCI images and propose reasonable system hardware modifications and the limits of certain image processing techniques.

X-ray diffraction (XRD) is proven to be an effective technique for baggage screening, as it can reveal inter- and intramolecular structural information of any solid substances (mostly polycrystalline), but also of liquids, aerosols and gels. The introduction of 2D pixelated energy resolved detectors, such as CdZnTe detectors, makes now possible the development of Energy Dispersive XRD (EDXRD) systems able to perform rapid in situ 3D baggage scanning. However, the EDXRD technique requires to fix the scattering angle to few degrees with very thick collimations, which induces lack of sensitivity and spatial resolution. It is then a question of proposing technological, architectural and algorithmic solutions to improve and find the best compromise between spatial resolution, power of discrimination and inspection speed. The CEA-LETI has designed CdZnTe based energy resolved detectors in which some specific detector-level signal processing are implemented to optimize both energy and spatial resolution thanks to subpixel positioning. Subpixel positioning enables to significantly improve both angular and spatial resolution in an EDXRD system. We implemented an EDXRD system using such a detection module coupled to a multiplexed secondary collimation, in which each physical pixel inspects the object at 4 or 5 different points. The sensitivity is more than four times better compared to a parallel collimation. Some specific iterative inversion algorithms reconstruct the diffraction signatures of the materials, even when they are close together as inside a baggage. Material discrimination performance limits were explored for several objects scenarios and various levels of photon statistics.

X-ray diffraction tomography (XDT) resolves the spatially-variant XRD profiles within the object, and provides improved material contrast compared to the conventional transmission-based computed tomography (CT). Due to the small diffraction cross-section, a typical full field-of-view XDT scan takes tens of hours using a table-top X-ray tube. In medical and industrial imaging applications, oftentimes only the XRD measurement within a region-of-interest (ROI) is required, which, together with the demand to reduce imaging time and radiation dose to the sample, motivates the development of interior XDT systems that scan and reconstruct only an internal region within the sample. However, existing interior reconstruction frameworks rely on a known region or piecewise constant constraint within the ROI, which do not apply to all the samples in XDT. In this presentation, we propose a quasi-interior XDT scheme that incorporates a small fraction of projection information from the exterior region to assist interior reconstruction. The low-resolution exterior projection data obviates the requirement for prior knowledge on the object, and allows the ROI reconstruction to be performed with the fast, widely-used filtered back-projection algorithm for easy integration into real-time XDT imaging modules. We also demonstrate the material classification based on the XDT profile reconstructed from pure interior and our combined ROI and exterior measurements.