Automatic compensation of phase aberration in digital holographic microscopy with deep neural networks for monitoring the morphological response of bone cells under fluid shear stress

In bone tissue, osteocytes are embedded within a microfluid-filled network which expose them to high levels of fluid shear stress (FSS). The osteocytes’ sensitivity to different levels of FSS has demonstrated. However, there are few attempts to image 3D cellular deformation under FSS by label-free and quantitative microscopy. Digital holographic (DH) microscopy is a powerful imaging technique that can provide rich intracellular information based on the refractive index (RI) contrast, without exogenous contrast agents. However, in DH image recording process, the recorded wave-front contains not only the object’s information but also the aberrations caused by the microscope objective (MO) and the imperfections of optical components of the system. The fitting-based numerical method removes total aberrations by detecting object-free background as reference surfaces. In this paper, we proposed a convolutional neural network (CNN) for multivariate regression to cope with the phase aberration compensation problem automatically thus allows performing long-term monitoring of bone cells morphological response under FSS. We transformed the problem of estimating the coefficients for fitting a phase aberration map to a regression problem. The aberrated phase images are put into this model which can automatically learns the internal features of phase aberrations. Then the optimal coefficients are estimated as an output of the network. Based on these coefficients, the phase aberration map is built by the polynomial fitting, and the phase aberrations are removed by subtracting the aberration phase image with the phase map. The trainning and validation set contain thousands of phase image of cells. The mean square error (MSE) is used as the loss function. Then, the trained model was used for aberrations compensation in the FFS experiment of osteocytes. The results show that the proposed approach can predict the optimal coefficients and automatically compensating the phase aberrations without detecting background regions and knowing any physical parameters.