传统的微生物培养和筛选技术普遍存在通量低、操作繁杂、耗费人力物力等问题。液滴微流控技术近十年来在微生物学研究、生物技术开发和与生物医疗产业中得到了广泛应用。微液滴具有通量高、体积小、易操作等特点，液滴微流控技术对人力和耗材的消耗更少，更易集成化和自动化，已日益成为微生物单细胞培养和分析的重要工具。在使用液滴微流控技术实现高通量培养和筛选微生物时，其中关键的一步是液滴分选。在皮纳升级微液滴系统中，开发细胞快速识别和实时检测技术在液滴分选中尤为重要。由于液滴流动速度快，这就要求检测系统具备响应迅速，灵敏度高，自动化和灵活等特点。目前已存在较为完备的检测和分选单细胞进行分析的方法，但大多数方法是有标记的，例如基于荧光信号或磁性标记。但由于有些细胞没有合适的生物标记物或标记具有潜在的细胞毒性，有些实验如干细胞分化等研究会受到标记的影响，因此有标记的方法存在局限性。其他如基于拉曼、光散射、核磁等无标记技术存在价格昂贵，搭建复杂等特点。针对上述问题，本文提出基于图像的高通量单细胞皮升级液滴微流控筛选系统，该系统与液滴微流控生成芯片与分选芯片结合，能实现高通量自动化的微生物单克隆的分离、培养、分选和收集，并开展相关应用示范。本研究结合基于图像的无标记成像和深度学习算法，实现了基于图像的高通量实时检测分选单克隆微生物。首先，搭建了相差显微系统，采用20倍物镜和帧率达700HZ高速相机，实现了更清晰更快速的无标记成像。通过生成芯片以高达1000HZ的速度稳定生成直径15-75μm的液滴。采用YOLOv8nano检测图像信号，识别准确率达98.7%，实现相机端采集图像、图像信息上传至模型、模型输出分选信号至控制端、微流控芯片下游通过介电泳实现分选，分选速率可达100HZ。其次，本研究完成了微生物单克隆的分离、培养、分选和收集全流程，证明了该系统的稳定性。综上所述，本研究提出了一种基于图像的高通量单细胞皮升级液滴微流控筛选系统，辅以深度学习算法实现实时高速检测和分选，为微生物高通量培养和分选提供了新的自动化平台，具有广泛的应用前景。
Droplet microfluidic technology has been widely applied in high-throughput microbial cultivation and screening over the past decade, due to its high throughput, ease of operation, lower resource consumption, easier integration and automation. In the picolitre droplet system, the development of rapid cell recognition and real-time detection technologies is particularly important in droplet sorting. While comprehensive methods exist for detecting and sorting single cells for analysis, most methods involve labeling, such as fluorescence signals or magnetic markers. However, labeled methods have limitations, as some cells may lack suitable biomarkers or labeling may have potential cytotoxicity or some experiment cannot use labeling. However, label-free techniques such as those based on Raman spectroscopy, light scattering, and nuclear magnetic resonance are expensive and complex to establish. With the continuous upgrades of cameras and computing hardware, and the iterative updates of deep learning algorithms, unlabeled detection methods based on bright field images have reached the threshold of high-throughput screening and have enormous research and application potential. In image-based unlabeled detection methods, efficiently tracking droplets and performing real-time image processing have always been a challenge. In recent years, object detection algorithms based on deep learning have developed rapidly, with inference times for single images compressed to tens of microseconds, which is much more efficient than traditional image processing methods, making real-time recognition, detection, and sorting of high-speed flowing droplets based on pictures possible. Therefore, this paper proposes an image-based high-throughput single-cell droplet microfluidics screening system, which is significantly optimized compared to existing image-based screening systems. For the first time, this system introduces an object detection algorithm and phase contrast imaging, and employs a cage structure, resulting in high throughput, label-free, miniaturization, automation, and low cost.This system, based on phase contrast microscopy imaging, uses the state-of-the-art You Only Look Once (YOLO) object detection algorithm and control module. In combination with the microfluidic chip, it can achieve high-throughput automated isolation, cultivation, sorting, and collection of microbial single clones. Firstly, a differential interference microscope system is constructed, employing a 20x objective lens and a high-speed camera with a frame rate of up to 700Hz to achieve clearer and faster label-free imaging. The generation chip reliably produces droplets with diameters ranging from 15 to 75 μm at speeds of up to 1000 Hz. The YOLOv8nano is used to detect image signals, achieving an identification accuracy of 98.7%. It enables image acquisition at the camera end, image information uploading to the model, model outputting sorting signals to the control end, and downstream microfluidic chip sorting via dielectrophoresis, with a sorting rate of up to 100Hz. Secondly, this study completes the entire process of isolation, cultivation, sorting, and collection of microbial single clones, demonstrating the stability of the system. In summary, this study proposes a high-throughput single-cell picolitre droplet microfluidic screening system based on imaging, supplemented by deep learning algorithms to achieve real-time high-speed detection and sorting. This system fills the gap of low throughput in image-based unlabeled screening systems and provides a new automated platform for high-throughput microbial cultivation and screening, with broad application prospects.
