The research question centers on improving the detection of threat materials, a critical aspect of security. Current methods often struggle with distinguishing between similar-appearing materials. This study investigates whether combining dark-field x-ray imaging with deep neural networks (DNNs) can overcome these limitations. Dark-field x-ray imaging, a phase-based method enhancing detail visibility by highlighting inhomogeneities below the spatial resolution of the system, offers a unique approach. It generates images displaying textural information specific to the material's microstructure. The hypothesis is that these textures, combined with conventional attenuation information, provide a richer dataset for improved material discrimination, especially when analyzed using the pattern recognition capabilities of DNNs. The study's importance lies in its potential to significantly enhance threat detection accuracy and efficiency, contributing to improved security measures.

Phase-based X-ray imaging, initially explored in the mid-1960s, gained traction in the mid-1990s due to advancements in synchrotron facilities. While initially focusing on enhanced contrast through phase effects in hybrid or phase-retrieved images, further research revealed the dark-field signal's potential. This signal arises from multiple refraction events caused by sub-resolution inhomogeneities, providing complementary information to attenuation and refraction. Various methods for dark-field extraction have been developed, including those using gratings, apertured masks (edge illumination or EI, used in this study), or crystals (at synchrotrons). Previous work demonstrated the information complementarity of these multi-modal imaging channels, prompting this investigation into leveraging this information using machine learning techniques.

The study utilizes an X-ray imaging system equipped with an X-Tek 160 source, a photon-counting detector (XCounter XC-FLITE FX2 CdTe CMOS), and custom-designed masks (pre-sample and detector masks). The edge illumination (EI) technique is employed for dark-field extraction, where illumination curves are analyzed to extract attenuation, dark-field, and refraction signals. A 4-way asymmetric mask enhances retrieval robustness by enabling an additional offset parameter. The data are processed to obtain attenuation, dark-field, and offset images at high and low x-ray energies. To address thickness dependence, ratios of thickness-dependent signals are calculated. Initial experiments using scatter plots visualize material discrimination, showcasing improved separation with combined attenuation and dark-field data. Two proof-of-concept CNN tests are performed. The first involves bags containing explosives or benign materials with clutter. Two CNN architectures (GoogleNet and Inception ResNet) are evaluated with various loss functions (softmax, cross-entropy, and hinge) and transfer learning from ImageNet. A second architecture incorporates texture recognition. The second test focuses on detecting C4 concealed in electronic items. A split-network architecture segments objects and then discriminates C4 textures. A de visu trial with security officers is compared against CNN performance. Data augmentation is investigated to evaluate its effect on the model performance. The performance of each CNN model is determined using the following metrics: Accuracy, Precision, and Recall.

The study's key findings demonstrate that combining dark-field and attenuation signals significantly enhances threat material discrimination. Using multiple x-ray energies further improves this discrimination by exploiting the energy-dependent differences in attenuation and phase effects. The effect of sample thickness is effectively mitigated by calculating ratios of thickness-dependent signals. Two proof-of-concept tests using CNNs show that the dark-field textures are well-suited for machine learning-based identification. In the first PoC test, using the Inception ResNet CNN architecture, no data augmentation, one additional fully connected layer for transfer learning, and hinge loss, along with the integration of texture recognition, achieved 99.8% accuracy in distinguishing between explosives and non-explosives. Removing the dark-field channel reduced accuracy to 93.6%. In the second PoC test, focused on detecting concealed C4 in electronic items, a split-network CNN achieved 100% true positive rate in a formal trial, significantly outperforming human security officers (48.8% true positives). Removing the dark-field data in this test reduced precision by 11% and overall accuracy by approximately 20%. Importantly, in both tests, the removal of dark-field data led to a marked degradation of performance, highlighting the crucial role of this signal in improving threat detection.

The findings strongly support the hypothesis that combining dark-field x-ray imaging and deep neural networks enhances threat material detection. The improved discrimination achieved by incorporating the dark-field signal, especially when using multiple energy levels, addresses a significant limitation of current dual-energy attenuation-only methods. The CNN models effectively learn to extract meaningful features from the complex textures generated by dark-field images, even when materials overlap, surpassing human performance in the second PoC test. The superior performance observed with the hinge loss function in the first test, compared to cross-entropy and softmax, further highlights the suitability of this specific loss function for addressing the subtlety of feature separation in this context. The success of the split-network architecture in the second test demonstrates its applicability to challenging scenarios where the target material is highly occluded.

This study demonstrates the significant potential of combining dark-field x-ray imaging and deep neural networks for enhanced threat material detection. The results from both proof-of-concept tests clearly show the superior performance of this combined approach compared to methods relying solely on multi-energy attenuation images. Future research should focus on expanding the dataset to include more materials, clutter, and variations in thickness, optimizing the CNN architecture based on the lessons learned here, and validating the approach in real-world settings. This technology holds promise for significantly improving security screening applications and may be applicable to various other fields requiring material discrimination based on microscopic structures.

The study's limitations include the relatively small size of the datasets used in both PoC tests and the heuristic approach to CNN architecture development. The limited datasets restricted ablation studies, especially in the first test, where k-fold cross-validation was employed instead of a separate held-out test set. The lack of data augmentation in the second test, while yielding good results, needs further investigation. Furthermore, the first PoC test used a single target material per sub-image, while the second test employed a limited set of materials and presented single-energy images to the human operators, which likely affected their performance. A larger scale study with more diverse and complex scenarios is required to fully assess the generalizability of the findings.