School of Electrical and Computer Engineering, Georgia Institute of Technology, Atlanta, GA 30332, USA.
Lab Chip. 2019 Oct 7;19(19):3292-3304. doi: 10.1039/c9lc00597h. Epub 2019 Sep 4.
Beyond their conventional use of counting and sizing particles, Coulter sensors can be used to spatially track suspended particles, with multiple sensors distributed over a microfluidic chip. Code-multiplexing of Coulter sensors allows such integration to be implemented with simple hardware but requires advanced signal processing to extract multi-dimensional information from the output waveform. In this work, we couple deep learning-based signal analysis with microfluidic code-multiplexed Coulter sensor networks. Specifically, we train convolutional neural networks to analyze Coulter waveforms not only to recognize certain sensor waveform patterns but also to resolve interferences among them. Our technology predicts the size, speed, and location of each detected particle. We show that the algorithm yields a >90% pattern recognition accuracy for distinguishing non-correlated waveform patterns at a processing speed that can potentially enable real-time microfluidic assays. Furthermore, once trained, the algorithm can readily be applied for processing electrical data from other microfluidic devices integrated with the same Coulter sensor network.
除了常规的粒子计数和粒径测量，库尔特传感器还可用于对悬浮粒子进行空间跟踪，在微流控芯片上分布多个传感器。库尔特传感器的代码复用允许通过简单的硬件实现这种集成，但需要先进的信号处理来从输出波形中提取多维信息。在这项工作中，我们将基于深度学习的信号分析与微流控码复用库尔特传感器网络相结合。具体来说，我们训练卷积神经网络来分析库尔特波形，不仅要识别某些传感器的波形模式，还要解决它们之间的干扰。我们的技术可以预测每个检测到的粒子的大小、速度和位置。我们表明，该算法在处理速度上能够实现 >90%的模式识别准确率，从而能够实现实时微流控分析。此外，一旦经过训练，该算法就可以轻松应用于处理与同一库尔特传感器网络集成的其他微流控设备的电数据。
