Label-free analysis of bacterial growth and lysis at the single-cell level using droplet microfluidics and object detection-oriented deep learning

Abstract

AbstractBacteria identification and counting at the small population scale is important to many applications in the food safety industry, the diagnostics of infectious diseases and the study and discovery of novel antimicrobial compounds. There is still a lack of easy to implement, fast and accurate methods to count populations of motile cells at the single-cell level. Here, we report a label-free method to count and localize bacterial cells freely swimming in microfluidic anchored picolitre droplets. We used the object detection oriented YOLOv4 deep learning framework for cell detection from bright-field images obtained with an automated Z-stack setup. The neural network was trained to recognizeEscherichia colicell morphology with an average precision of approximately 84%. This allowed us to accurately identify individual cell division events, enabling the study of stochastic bacterial growth starting from initial populations as low as one cell. This work also demonstrates the ability to study single cell lysis in the presence of T7 lytic bacterial viruses (phages). The high precision in cell numbers facilitated the visualization of bacteria-phage interactions over timescale of hours, paving the way towards deciphering phage life cycles in confined environments.