Lab-based 3D x-ray microscopy has advanced in capability and popularity over the past decade due to the unique non-destructive high-resolution 3D imaging capabilities and the concurrent availability of advanced commercial instruments. However, especially for the highest-resolution imaging tasks, laboratory x-ray microscopy has traditionally suffered from lower image quality and/or longer scan times as compared to similar synchrotron capabilities. Recent progress in 3D image reconstruction has brought about dramatic improvements in achievable scan speed and image quality, irrespective of advancements in x-ray source or detector technologies. These advancements have driven x-ray CT image quality and throughput beyond what is achievable through traditional analytical methods like the Feldkamp-Davis-Kress (FDK) algorithm [ 1], with many of these new reconstruction methods available in commercial products.
Reconstruction algorithms leveraging deep learning have emerged as promising methods for 3D x-ray microscopy image reconstruction with reduced artifacts and increased signal to noise ratio. In this paper we present a novel technique for image restoration using a synthetic prior intermediate (SPI), with application to noise and artefact removal in X-ray reconstruction. In SPI based restoration an SPI, constructed using existing state-of-the-art techniques, is forward modelled using physics or heuristic techniques to create a synthetic data estimate. A neural network based inverse restoration technique is then optimized to remove the artefacts introduced by this forward model. This restoration technique is then applied to the original data, creating a high-quality final reconstruction.
We apply this approach to noise and sampling artefact removal, using Deep Learning based reconstruction to construct the SPI, then a heuristic data-driven model to create the synthetic data estimate.
Smart Citations