Automation of manipulating and transferring micro-scale particles using deep learning

Abstract

This thesis describes a study of automation of manipulating and transferring micro-scale particles using deep learning for intelligent cell sorting and efficient sample delivery methods. To this end, we propose an integrated cell sorting system, an automated sample delivery device, and key technologies such as cell image analysis, a TensorRT-optimized deep learning model, and efficient automated sample delivery methods. In the proposed integrated system, the deep learning model was optimized under the TensorRT framework to enable real-time cell analysis and sorting. The automated sample delivery device efficiently delivered a small volume of micro-scale crystals for serial femtosecond crystallography (SFX). This thesis consists of a study on micro-scale particle manipulation using deep learning and a study on efficient micro-scale particle transfer method. In the particle manipulation part, a deep learning-based micro-scale particle sorting system is proposed. To enable the high-speed analysis of fast-flowing cells in a microfluidic channel, we implemented multi-thread parallel computations: camera image acquisition, cell detection, cell classification, and cell sorting. The inference time of the deep learning model was also reduced by 4-5 times by converting the trained model into a TensorRT-optimized deep learning model. In addition, the syringe-driven piezo sorting device was used for the first time in order to simplify the experimental procedures and the chip manufacturing process. The performance of the deep learning-based micro-scale particle sorting system was verified through the real-time sorting of micro-scale beads and cells. The target beads and cells were successfully separated with high purities from the bead and cell mixtures (15 μm* & 10 μm beads: 98.0%: HL-60* & Jurkat cells: 95.1%: HL-60* & K562 cells: 94.2%, * is target beads or cells). In addition, the particle sorting system was upgraded by adding the following features: (1) a vertically standing syringe pump to prevent particle sedimentation, (2) an LED strobe light to obtain blur-free images of fast-flowing cells, (3) a high-precision piezo stage to obtain in-focus images of linearly aligned flowing cells in the microchannel. The performance of the upgraded system was verified by the real-time sorting of micro-scale beads, and the target beads (15 μm) were successfully separated from the bead mixture (15 μm & 10 μm) with a high sorting purity of 99.4%. In the particle transfer part, an automated sample delivery device for micro-scale crystals is proposed. Serial femtosecond crystallography can determine the structure of protein crystals by analyzing the diffraction patterns obtained by irradiating the crystals with an X-ray free electron laser (XFEL). To determine the structure of protein crystals purified in a small volume, it is important to develop an efficient sample delivery method for SFX experiments. Thus, a combined inject-and-transfer system (BITS) that can efficiently deliver protein crystals to the XFEL interrogation point is proposed. The BITS allows for solution samples to be reliably deposited on ultraviolet ozone (UVO)-treated polyimide films, at a minimum flow rate of 0.5 nl/min, in both vertical and horizontal scanning modes. To demonstrate the utility of BITS in SFX experiments, lysozyme crystal samples were embedded in a viscous lard medium and injected at flow rates of 50–100 nl/min through a syringe needle onto a UVO-treated polyimide film. Using the BITS method, the room-temperature structure of lysozyme was successfully determined at a resolution of 2.1 Å. The BITS system was upgraded by installing a high-precision piezo stage for real-time motion control of the injection needle and developing a graphical user interface (GUI) for user convenience. Moreover, the development of an inject-and-diffuse method for time-resolved studies with liquid applications in the BITS and its preliminary results are reported.