Multi-contrast x-ray identification of inhomogeneous materials and their discrimination through deep learning approaches

Abstract

Recent innovations in x-ray technology (namely phase-based and energy-resolved imaging) offer unprecedented opportunities for material discrimination; however, they are often used in isolation or in limited combinations. Here we show that the optimized combination of contrast channels (attenuation at three x-ray energies, ultra-small angle scattering at two, standard deviation of refraction) significantly enhances material identification abilities compared to dual-energy x-ray imaging alone, and that a combination of off-the-shelf machine learning approaches can effectively discriminate, e.g., threat materials, in complex datasets. The methodology is validated on a range of materials and image datasets that are both an order of magnitude larger than those used in previous studies. Our results can provide an effective methodology to discriminate, and in some cases identify, different materials in complex imaging scenarios, with prospective applications across the life and physical sciences. While the detection of threat materials is used as a demonstrator here, the methodology could be equally applied to, e.g., the distinction between diseased and healthy tissues or degraded vs. pristine materials.