Deep learning-based accurate and rapid tracking of 3D positional information of microparticles using digital holographic microscopy

Abstract: Digital holographic microscopy (DHM) encrypts three-dimensional (3D) volumetric information of a test sample in the form of diffraction patterns. Computationally intensive numerical reconstruction and autofocusing procedures, which require precise determination of some parameters, are essential for the extraction of 3D positional information. In this study, a method that can satisfy these requirements by combining digital in-line holographic microscopy (DIHM) with deep learning was proposed. Deep neural networks were trained with tens of thousands of defocused holograms of microparticles moving in a microtube. The in-plane positions of the spherical microparticles were detected by applying Segnet and circular Hough transform, and the depth positions of the microparticles were estimated using a convolutional neural network. The performance of the proposed method was verified by conducting a planar surface experiment and tracking the 3D motions of the particles in a circular microtube. Results indicated that the trained neural networks precisely determined the 3D positions and 3D trajectories of the particles, and the proposed method outperformed conventional methods in measuring 3D positional information in the in-plane and out-of-plane direction. The process time of the proposed method for 3D tracking was only 1.92% of that of the conventional method. The present deep learning-based DIHM method can facilitate the analysis of the 3D dynamic motions of particles or cells and 3D tracking of numerous samples in motion. Graphic abstract: [Figure not available: see fulltext.].