Performance of autofocus capability of deep convolutional neural networks in digital holographic microscopy

Digital holographic microscopy (DHM) is a label-free, single-shot technique that is well suited for imaging living three dimensional samples [1]. DHM overcomes a problem present in conventional optical microscopes of a shallow depth-of-field, permitting one to reconstruct at different in-focus planes of a volume. Nevertheless, an object of interest is usually in-focus only in one or few depths as a single reconstruction layer still carries a shallow depth-of-field.Different methods to find objects in focus have been proposed. These methods can be based on self-entropy [2],amplitude analysis [3], power spectra [4] or other metrics [5]. Common to all of these single wavelength methods is that they reconstruct a stack of images that the focus metric is applied to. This procedure is applied to each of the holograms. We propose to use deep learning as an autofocusing method. The greatest benefit of the proposed method is that after the training is completed, the in-focus plane can be obtained by using only the single in the hologram plane and without any reconstruction