Enhanced material discrimination with dark field imaging combined with deep learning

Dark field imaging relies on the detection of x-rays refracted in multiple directions by features smaller than the spatial resolution of the imaging system. As such it is sensitive to the microstructure of a material, and spatial variations of this microstructure create a texture in the dark field image that can be unique to that material. This is a perfect match with deep learning algorithms, which can be trained to recognise specific textures. We applied this approach to the detection of explosive materials, creating realistic, challenging scenarios where these were hidden in small quantities in electrical items placed inside bags alongside other “cluttering” objects, mimicking the content of a carry-on bag, and obtained a 100% detection rate. While this result is surely affected by the small scale of our study and by the fact it was carried out in a relevant, but not an operational environment, it indicates a significant potential for this new “combined” approach. Importantly, there is nothing special about explosives as such, and the method can be applied to any microscopically inhomogeneous material.

Examples images from the study, specifically of a hair drier (top) and a mobile phone (bottom), with conventional attenuation-based x-ray images (centre) directly compared to their dark-field correspondent (right).