The effect evaluation of the antibiotic susceptibility test based on bacterial solution is of great significance for clinical diagnosis and prevention of antibiotic abuse. Applying a microfluidic chip as the detection platform, the detection method of using microscopic images to observe bacteria under antibiotic can greatly speed up the detection time, which is more suitable for high-throughput detection. However, due to the influence of the depth of the microchannel, there are multiple layers of bacteria under the focal depth of the microscope, which greatly affects the counting and recognition accuracy and increases the difficulty of relocation of the target bacteria, as well as extracting the characteristics of bacterial liquid changes under the action of antibiotics. After the focal depth of the target bacteria is determined, although the z-axis can be controlled with the help of a three-dimensional micro-operator, the equipment is difficult to operate and the long-term changes of the target bacteria cannot be tracked quickly and accurately. In this paper, the YOLOv5 algorithm is adopted to accurately identify bacteria with different focusing states of multi-layer bacteria at the z-axis with any focal depth. In the meantime, a certain amount of microspheres were mixed into bacteria to assist in locating bacteria, which was convenient for tracking the growth state of bacteria over a long period, and the recognition rates of both bacteria and microspheres were high. The recognition accuracy and counting accuracy of bacteria are 0.734 and 0.714, and the two recognition rates of microspheres are 0.910 and 0.927, respectively, which are much higher than the counting accuracy of 0.142 for bacteria and 0.781 for microspheres with the method of enhanced depth of field (EDF method). Moreover, during long-term bacterial tracking and detection, target bacteria at multiple z-axis focal depth positions can be recorded by the aid of microspheres as a positioning aid for 3D reconstruction, and the focal depth positions can be repositioned within 3-10 h. The structural similarity (SSIM) of microscopic image structure differences at the same focal depth fluctuates between 0.960 and 0.975 at different times, and the root-mean-square error (RMSE) fluctuates between 8 and 12, which indicates that the method also has good relocation accuracy. Thus, this method provides the basis for rapid, high-throughput, and long-term analysis of microscopic changes (e.g., morphology, size) of bacteria detection under the addition of antibiotics with different concentrations based on microfluidic channels in the future.
Keywords: 3D relocation; deep learning; detection of antibiotic susceptibility; identification of bacteria; microfluidics.