High-throughput techniques for tracking bacteria in 3D

Digital holographic microscopy is a broadly used technique for tracking cells in three dimensions. This methodology consists of the recording of holograms using a special experimental setup designed to record the diffracted light of a weakly scattering sample illuminated by a coherent light source. Localisation in three dimensions is done by refocusing the recorded images at different axial positions to then apply a Sobel-type kernel to localise the cells by looking into intensity changes in the axial direction. The next step of the process is to detect individual trajectories using a coordinate-stitching algorithm, usually a “search sphere” algorithm. The entire process can be computationally expensive. This work presents alternatives to the steps of this process in order to improve its throughput. First, a modified propagator is presented as an alternative to using the Rayleigh-Sommerfeld/Sobel scheme to localise Archaea, strain Haloarcula (HGSL) in three dimensions by combining both steps of the process into just one. It is shown that the modified propagator is computationally faster by up to 21%, while maintaining axial precision that lies within 1.75 μm from Rayleigh-Sommerfeld/Sobel scheme. Track metrics such as track mean speed and angle change also lie within 1 μms−1 and 1◦ respectively from the values obtained the Rayleigh-Sommerfeld/Sobel localisation. Also, an alternative localisation technique based on image cross-correlation is presented to track Escherichia coli. Two alternatives, GPU and Optical, for computing the image cross-correlations are discussed. The GPU approach uses a Graphical Processing Unit to perform the calculations while the optical approach utilises a sophisticated system based on a 4-f optical correlator to compute image cross-correlations using light. These GPU and Optical approaches have been found to be faster in localising cells than the usual Rayleigh-Sommerfeld/sobel scheme by up to 70% and 1000% respectively. Localisation with GPU lies within 2.5 μm while optical localisation lies within 7.3 μm with respect to the usual method. Finally, cell tracking using density-based spectral clustering of applications with noise (DBSCAN) machine learning clustering algorithm is compared against the usual search sphere algorithm. It is shown that this proposed algorithm is up to 57.76 times faster than the search sphere algorithm. Track metrics such as track mean speed and angle change also lie within 0.002 μms−1 and 0.06◦ respectively from the values calculated from search sphere detected tracks. The proposed methods for particle localisation using image cross correlation calculated with a GPU or an ”optical computer”, and for bacterial 3D tracking using machine learning, present advances in achieving the desired results more quickly than the methods currently in use.