Deep learning-assisted wavefront correction with sparse data for holographic tomography

In this paper, a novel approach using deep learning-assisted wavefront correction in beam rotation holographic tomography to acquire three-dimensional images of native biological cell samples is described. With digitally recorded holograms, the wavefront aberration is contained in the reconstructed phases. However, there are large computation costs for compensating the phase aberration during the reconstruction. With the aid of a deep convolution network, we present an effective algorithm on the reconstructed phases with sparse data for active wavefront correction. To accomplish this, we developed a Res-Unet scheme to segment the cell region from its background aberration and a deep regression network for the representation of the aberration on Zernike orthonormal basis. Moreover, a sparse data fitting algorithm was used to predict the Zernike coefficients of whole scanning angles from the collected sparse data. As a result, the proposed algorithm is capable of accurately correcting the background aberration and much faster than the original plain algorithm.