Phase-aberration compensation via deep learning in digital holographic microscopy

Abstract

Digital holographic microscopy (DHM), a quantitative phase-imaging technology, has been widely used in various applications. Phase-aberration compensation in off-axis DHM is vital to reconstruct topographical images with high precision, especially for microstructures with a small background or a dense phase distribution. We propose a numerical method based on deep learning combined with DHM. First, a convolutional neural network (CNN) recognizes and segments the sample and the background area of the hologram. Zernike polynomial fitting is then executed on the extracted background area. Finally, the whole process of phase-aberration compensation is automatically performed. To obtain a robust and accurate deep-learning model for hologram segmentation, we collected many holograms corresponding to several samples that had different morphological characteristics. The experimental results confirm that the trained CNN can accurately segment the sample from the background area of the hologram, and that this method is applicable and effective in off-axis DHM.

Advances in Digital Holographic Interferometry

“…The passage of images through such blocks trains the neural network by the method of backpropagation. Thus, neural networks solve the problems of hologram filtering and reconstruction and mapping depth [ 51 ], as well as eliminating phase aberrations [ 52 ].…”

“…The passage of images through such blocks trains the neural network by the method of backpropagation. Thus, neural networks solve the problems of hologram filtering and reconstruction and mapping depth [51], as well as eliminating phase aberrations [52]. The optical thickness distribution 𝐿 can finally be acquired with the suppressed amplified noises and the free twin-image [51].…”

Synthetic-wavelength phase guided phase unwrapping method for dual-wavelength digital holography

“…The complex wavefront of object beam could be obtained by Fourier transform method and angular diffraction theory. [26][27][28] Then the wrapped phase of the complex wavefront could be obtained, which carries the height information of the object to be measured. After subtraction of two phase maps, a larger measurement range can be obtained, but more severe noise is introduced.…”

“…In the off-axis dual-wavelength DH, the digital holograms of two wavelengths are recorded respectively by digital recording device charge coupled device (CCD) or complementary metal oxide semiconductor. The complex wavefront of object beam could be obtained by Fourier transform method and angular diffraction theory 26 – 28 Then the wrapped phase of the complex wavefront could be obtained, which carries the height information of the object to be measured.…”

Machine Learning Assisted Classification of Cell Lines and Cell States on Quantitative Phase Images

“…Machine-learning approach can be successfully used in several routine tasks of data processing in digital holography, for example in aberration compensation [11], assistance in increase of spatial resolution by automatic stitching of data sets in synthetic aperture digital holography [12], tomographic reconstruction of three-dimensional distributions of refractive index [13] and identification of intracellular compartments during analysis of such distributions [14].…”