Towards Real-Time Holographic Three-Dimensional Imaging with Machine Learning

Abstract

Real-time visualization of cellular and molecular dynamics is critical in achieving a flexible, adaptive imaging platform capable of automatically identifying regions of interest and proactively modifying acquisition parameters In holographic microscopy, the time-intensive steps of sample and / or stage manipulation are replaced by the computational time required to accomplish the volume reconstruction. A key challenge in real time visualisation of time-lapse holographic microscopy arises when long-term studies are underway, such as the study of biofilm development, which may take place over hours or days. In this case, computing 3-D positions and reconstructions is a CPU-intensive process. To address this, efforts are underway to apply GP-GPU based approaches that allow for parallelization of the process. In the absence of GPU platform, an attractive alternative is to consider the use of artificial neural networks(ANN) to bypass the repeated matrix multiplication associated with each reconstructing distance. Here, we demonstrate how supervised ANN can be used by determining the algorithms necessary to estimate the z-position of particles from a single 2-dimensional in-line hologram. With only a small cost of one-time training, the newly-found algorithm would bypass the iterative calculations required to search through the dataset to identify the final position of objects in each newly acquired hologram. This in turn will help realise real-time visualisation of holographic imaging and achieving a adaptive imaging system where executive decisions on the speed and volume of imaging interest can be made during acquisition.