Improving Image Reconstruction for Ultra-Fast Ptychographic Acquisitions via Deep Learning Denoising

Abstract X-ray ptychography is a scanning coherent diffraction imaging technique which combines nanometer-scale resolution with high penetration depth. This method has been proven to be suitable for scanning weakly absorbing samples and therefore potentially very valuable for medical applications such as brain imaging. However, currently employed scanning techniques present challenges: step-scanning is too slow and inefficient, while fly-scanning introduces blurring and noise into reconstructions due to the motion and reduced photon counts per pixel. To date, only a few methods have been proposed to denoise reconstructions, most of which rely on traditional approaches and are limited in addressing the challenges posed by noise and blurring. To overcome these limitations, we investigate the possibility of using a deep learning-based denoising method combined with position binning. The deep learning-based denoising method, Deep Image Prior (DIP), denoises the reconstructions while position binning increases the photon count statistics per pixel. The method can be integrated within the existing iterative phase retrieval algorithms to denoise the object or probe in between iterations. The method is tested in far-field geometry on two different samples: a Siemens star resolution target and a polymer-based phantom mimicking the white matter of the brain. By assessing the resolution via Fourier ring correlation, we measure up to a 14% increase in the resolution. However, depending on the architecture used, artifacts due to machine hallucination appear in the denoised images which could be affecting the observed enhancement in resolution. This will be the subject of further investigation.