High-throughput, sustained bacterium tracking in 3D with digital holographic microscopy integrated with intensity-based detection, deep learning, and cross-correlation

Temporally extended, three-dimensional trajectories of microbes are particularly useful to understand bacterial motility behaviors and swimming patterns. Conventional particle detection methods employed in digital holographic microscopy (DHM), e.g., intensity-based methods, are often ineffective in producing long trajectory data. We develop a novel hybrid particle detection and tracking method that can extract longer trajectories in 3D with higher throughput and fidelity. This method integrates conventional intensity-based particle detection, deep learning image classification, and cross-correlation displacement determination. The conventional detection provides seeds (in terms of 3D coordinates) to the cross-correlation operation, which is implemented on two sequential light fields reconstructed from consecutive holograms. Peak detection on the resultant correlation map gives a 3D vector that points to the same particle in the next frame. This realizes sustained particle detection and tracking in one step. Additionally, a trained deep learning classifier is used to differentiate between true and false particle detections to eliminate false pairings. For demonstration, the new method was applied to experimental holograms of bacteria and has produced significantly more long trajectories that are all valid detections. In addition, the measurement accuracy in the depth direction is also superior to the conventional method.