作者：俞朱恺1, 陈明灿1, 许婧靓1, 王驰2
作者单位：1. 上海大学中欧工程技术学院, 上海 200444;
2. 上海大学机电工程与自动化学院, 上海 200444
关键词：微流控;循环肿瘤细胞;深度学习;目标检测;YOLO算法
摘要：
目前，医学检测可用的技术大多依赖于针对细胞生物学特性的分离，存在灵敏度低、技术程序复杂、成本效益低、不适合连续监测等问题。因此，该文提出一种基于深度学习的微流控新技术，用于模拟血流环境中进行循环肿瘤细胞的精准检测。为辅助检测框定位，该文使用有限元分析软件模拟微流体液滴在不同流量水相（分散相）与油相（连续相）中的产生过程，并据此选择最稳定流量（分散相：1 μL/min，连续相：16 μL/min）。为验证模拟结果的准确性，该文在不同流量的视频中进行定位测试，通过检测框的像素和尺度优化将相对误差控制在<1%。在此基础上，通过加入注意力机制和多尺度特征融合算法模块对YOLOv5算法结构进行升级，进而实现液滴的精确定位和尺寸预测。实验中，该文分别将肺癌和乳腺癌细胞载入水相，以此构建数据集（15 min，20 F/s）进行算法模型训练。最终，改进后的YOLO微全分析系统可以准确地测量液滴的位置（平均精度均值：98.92%）和大小(相对误差：0.49%)，并精准识别视频流中的小目标循环肿瘤细胞（准确率：72.49%）。该文不仅为微流体液滴中成分的智能检测提供了新技术，而且有望实现真实血液环境中循环肿瘤细胞的实时监测。
YOLO micro-analysis system for accurate detection ofcirculating tumor cells
YU Zhukai, CHEN Mingcan, XU Jingjing, WANG Chi
1. Sino-European School of Technology of Shanghai University, Shanghai University, Shanghai 200444, China;
2. School of Mechatronic Engineering and Automation, Shanghai University, Shanghai 200444, China
Abstract: Currently, most of the available technologies for medical diagnosis rely on separation based on cell biological properties, which suffer from problems such as low sensitivity, complex technical procedures, low cost-effectiveness, and unsuitability for continuous monitoring. Therefore, this work proposes a new microfluidic technology based on deep learning to adapt to the accurate detection of circulating tumor cells in a high-flow environment. To assist in positioning the detection frame, a finite element analysis software was used to simulate the generation process of microfluidic droplets in the water phase (dispersion phase) and oil phase (continuous phase) at different flow rates, so as to select the most stable flow rate (dispersion phaseat 1 μL/min and continuous phase at 16 μL/min, accordingly). To verify the results of simulation, positioning tests were conducted on videos with different flow rates, and the relative error was controlled within 1% via the pixel and scale optimization of the detection frame. On this basis, the YOLOv5 algorithm structure was optimized by adding an attention mechanism and a multi-scale feature fusion algorithm module, converting the pixels and scales of the detection frame, thereby achieving precise droplet detection and size prediction. In the experiment, lung cancer and breast cancer cells were added into the water phase to construct a data set (15 min, 20 F/s) for algorithm model training. Finally, the improved YOLO micro-total analysis system could accurately measure the position (mean average precision of 98.92%) and size (relative error of 0.49%) of the droplets, and precisely identify circulating tumor cells in the video stream with an accuracy of 72.49%.This work not only provides a new technology for the intelligent detection of components in microfluidic droplets, but also provides a potential strategy for real-time monitoring of circulating tumor cells in real blood environment.
Keywords: microfluidics;circulating tumor cells;deep learning;target detection;YOLO algorithm
2023, 49(11):1-6,15 收稿日期: 2023-09-30;收到修改稿日期: 2023-10-29
基金项目: 国家自然科学基金 (22001162，62175144)
作者简介: 俞朱恺(2000-),男,浙江嘉兴市人,硕士研究生,专业方向为医学图像和生物信息。
参考文献
[1] DONG Q, LIC X, ZHAO L, et al. Advances in forensic DNA rapid test[J]. China Measurement & Test, 2017, 43(7): 59-65.
[2] HAN J P, SUN J, OU Y, et al. Application of microfluidics technology in DNA rapid testing of forensic science[J]. China Measurement & Test, 2016, 42(1): 53-60.
[3] REDMON J, DIVVALA S, GIRSHICK R, et al. You only look once: unified, real-time object detection[C]//Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2016.
[4] TAN M, PANG R, LE Q V. Efficientdet: scalable and efficient object detection[C]//Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2020.
[5] ZHANG S, LIANG X, HUANG X, et al. Precise and fast microdroplet size distribution measurement using deep learning[J]. Chemical Engineering Science, 2022, 247: 116926.
[6] HE K, GKIOXARI G, DOLLÁR P, et al. Mask r-cnn[C]//Proceedings of the IEEE International Conference on Computer Vision, 2017.
[7] HOWELL L, ANAGNOSTIDIS V, GIELEN F. Multi‐object detector yolov4‐tiny enables high‐throughput combinatorial and spatially‐resolved sorting of cells in microdroplets[J]. Advanced Materials Technologies, 2022, 7(5): 2101053.
[8] DURVE M, BONACCORSO F, MONTESSORI A, et al. A fast and efficient deep learning procedure for tracking droplet motion in dense microfluidic emulsions[J]. Philosophical Transactions of the Royal Society A, 2021, 379(2208): 20200400.
[9] YONG Y M, YANG C, JIANG Y, et al. Numerical simulation of immiscible liquid-liquid flow in microchannels using Lattice Boltzmann method[J]. Science China Chemistry, 2011, 54: 244-256.
[10] PENG L, YANG M, GUO S, et al. The effect of interfacial tension on droplet formation in flow-focusing microfluidic device[J]. Biomedical microdevices, 2011, 13: 559-564.
[11] LI X B, LI F C, YANG J C, et al. Study on the mechanism of droplet formation in T-junction microchannel[J]. Chemical Engineering Science, 2012, 69(1): 340-351.
[12] JANGIR P, JANA A K. CFD simulation of droplet splitting at microfluidic T-junctions in oil–water two-phase flow using conservative level set method[J]. Journal of the Brazilian Society of Mechanical Sciences and Engineering, 2019, 41: 1-16.
[13] OSHER S, SETHIAN J A. Fronts propagating with curvature-dependent speed: algorithms based on Hamilton-Jacobi formulations[J]. Journal of Computational Physics, 1988, 79(1): 12-49.
[14] JOCHER G, CHAURASIA A, STOKEN A, et al. Ultralytics/yolov5: v6. 1-TensorRT, TensorFlow edge TPU and OpenVINO export and inference[Z]. Zenodo, 2022.
