Wrapped phase aberration compensation using deep learning in digital holographic microscopy

In digital holographic microscopy (DHM), phase aberration compensation is a general problem for improving the accuracy of quantitative phase measurement. Current phase aberration compensation methods mainly focus on the continuous phase map after performing the phase filtering and unwrapping to the wrapped phase map. However, for the wrapped phase map, when larger phase aberrations make the fringes too dense or make the noise frequency features indistinct, either spatial-domain or frequency-domain based filtering methods might be less effective, resulting in phase unwrapping anomalies and inaccurate aberration compensation. In order to solve this problem, we propose and design a strategy to advance the phase aberration compensation to the wrapped phase map with deep learning. As the phase aberration in DHM can be characterized by the Zernike coefficients, CNN (Convolutional Neural Network) is trained by using massive simulated wrapped phase maps as network inputs and their corresponding Zernike coefficients as labels. Then the trained CNN is used to directly extract the Zernike coefficients and compensate the phase aberration of the wrapped phase before phase filtering and unwrapping. The simulation results of different phase aberrations and noise levels and measurement results of MEMS chip and biological tissue samples show that, compared with current algorithms that perform phase aberration compensation after phase unwrapping, the proposed method can extract the Zernike coefficients more accurately, improve the phase data quality of the consequent phase filtering greatly, and achieve more accurate and reliable sample profile reconstruction. This phase aberration compensation strategy for the wrapped phase will have great potential in the applications of DHM quantitative phase imaging.