Fast automated quantitative phase reconstruction in digital holography with unsupervised deep learning

Abstract

Digital holography can provide quantitative phase images related to the morphology and content of biological samples. To reconstruct an accurate phase image, several processes, such as phase unwrapping, focusing, and calculation of digital reference wave and numerical propagation, are essential. However, this process is time-consuming. We propose a model that performs phase reconstruction in one-step and two-step using an unsupervised image-to-image translation structure. The two-step reconstruction model translates the phase image, which is obtained by performing numerical propagation on the hologram, into an accurate phase image, whereas the one-step reconstruction model directly translates the hologram into an accurate phase image. The proposed model shows similar high-performance reconstruction to the supervised learning model used in many previous studies. However, since supervised learning is trained in strict pairs, many target domain data (accurate phase imagery) is required. Since the proposed model is trained by unsupervised learning, phase reconstruction can be performed with a small amount of target domain data. The proposed method can help to observe the morphology and movement of biological cells in real-time applications.