Towards Practical X-ray Microscopy via Optimization and Deep Learning
X-ray diffraction microscopy is an emerging technology with the capability of providing high-quality imaging of nanoscale specimens such as viruses, proteins, and crystals. While very promising, the development of this technology is challenged by the fact that measurements are often highly corrupted by both noise and missing low-frequency data. This problem is especially challenging when seeking to image fragile biological specimens which can only safely tolerate low-radiation photon fluxes, since the levels of noise corruption are inversely proportional to X-ray photon flux values. To overcome these challenges, this work introduces new physical and algorithmic methods for X-ray microscopy which enable imaging at far improved resolutions, even when data is collected via low-radiation measurements. Specifically, this is achieved via the physical implementation of a holographic reference object, i.e. holographic X-ray microscopy, as well as the employing of noise-aware image reconstruction algorithms. These algorithms consist of both classical optimization and deep learning-based methods which incorporate the noise models encountered in X-ray diffraction microscopy into the image reconstruction pipeline.