New technologies for x-ray microscopy: phase correction and fully automated deep learning based tomographic reconstruction

Abstract

We introduce two new technologies which significantly improve X-ray microscopy image quality and contrast through the removal of spurious or undesired reconstructed signal. First, we introduce the first technique for the fully automated training of deep learning networks for X-ray reconstruction – DeepRecon Pro. These networks can remove or greatly reduce two major contributors to spurious reconstructed image signal; random noise and sparse sampling artefacts. These two sources of spurious signal are discussed, showing how they can be differentiated, characterized, and removed with correctly trained reconstruction networks. This is showcased on a synthetic rock sample high in spurious image signal. Both the pure noise field and the sparse sampling artefact field were calculated, with the pure noise field being approximately pixelwise random, whereas the artefact field had a much larger characteristic spatial wavelength (the precise structure of this field is highly sample dependent). We then show the performance of this technology across a range of datasets, all of which show a robust and consistent loss pattern, and greatly reduced image noise and artefact levels. We also test network performance on a series of datasets outside the network’s training region, showing a consistent high level of performance. Finally, we introduce a technique which allows for the complete removal of propagation phase contrast artefacts which can become the dominant contrast mechanism when imaging at high resolution or at low kV. The removal of this effect reveals the inherent reconstructed material contrast which can then be much more effectively denoised, segmented or analyzed.