Wednesday, August 4 | 4:00 pm EDT

Next generation XRM imaging with DeepRecon Pro and PhaseEvolve


Matt Andrew, Ph.D.

Carl Zeiss Microscopy GmbH


X-ray microscopy has grown to be the leading method for non-invasive 3D microstructural characterization of samples ranging from electronics to materials sciences to rocks, but while much progress has been made, many challenges remain. One of the biggest challenges associated with X-ray microscopy is that of speed. Laboratory photon sources are relatively weak when compared to national synchrotron facilities. This leads to a fundamental tradeoff between imaging throughput and image quality. The past 10 years have led to a revolution in the development and application of a broad range of technologies broadly grouped together under the umbrella of “machine learning” or artificial intelligence, however its deep application to the field of micro-structural characterization is still in its early stages. Perhaps the biggest barrier to adoption of these technologies has been the requirement for deep level of expertise in machine learning to harness and apply the technology to a specific problem. In this presentation we introduce DeepRecon Pro, a reconstruction technology that allows for the power of deep learning to be brought into every laboratory, no matter the researcher’s familiarity with machine learning techniques. Its unparalleled ease-of-use allows for models to be trained on X-ray datasets with a single click, effectively improving throughput by up to an order of magnitude (up to 10x), or greatly enhancing image quality (reducing noise or imaging artefacts) at a fixed throughput. We will showcase the qualitative and quantitative application of this technology on a range of sample types. When imaging batteries, high resolutions must be achieved within entire packaged samples as disassembling samples is dangerous and technically challenging, leading to long acquisition times and relatively noisy images. This noise can overprint the extremely subtle contrast exhibited by the battery’s graphite anode, impregnated with electrolyte. DeepRecon Pro can be used to greatly reduce the noise floor in the image, enabling the subtle anode contrast to be seen. Models trained to improve imaging throughput allow for cathode material to be imaged 4x faster without degrading image quality. We will also use quantitative pore size distribution analysis to show DeepRecon Pro recovers critical sample defects in an inconel lattice which would otherwise be lost when reducing scan times.

Another significant challenge in X-ray imaging are artefacts arising from the complicated Multiphysics of X-ray interaction. Image reconstruction typically assumes that the only significant photon interaction is the photo-electric effect, leading to Beer-Lambert linear attenuation. While generally giving good results, when imaging at high resolution, or when imaging at low energies propagation phase contrast effects become a significant contributor to total X-ray interaction. While these effects can be useful for accentuating edges of interfaces, they can interfere with segmentation and quantitative sample assessment. In this presentation we will introduce PhaseEvolve - a newly available technique which removes phase effects using an analytical description of the associated artefact, allowing for much easier and more accurate segmentation. This is shown on pharmaceutical samples, allowing for greatly increased contrast and material differentiation. It is also shown on an aluminum powder sample for additive manufacturing, allowing accurate segmentation and included porosity analysis.