Development of a robust segmentation network for single-cell impedance cytometry data

Microfluidic impedance cytometry (MIC) is a high-throughput, label-free technology for analysing individual flowing cells by examining their interactions with a multifrequency electric field. This study investigates the integration of machine learning (ML) with MIC to enable effective processing of raw impedance data streams (i.e., electric current signals). Specifically, the research highlights the robustness and adaptability of neural networks for signal segmentation (i.e., event detection). To this aim, impedance data from various microfluidic chip designs were considered, which are characterized by event-signals with different temporal shapes, including asymmetric and symmetric bipolar signal profiles. Four neural network architectures, Bidirectional Long Short-Term Memory (biLSTM), Gated Recurrent Units (GRU), Convolutional Neural Networks (CNN), and Temporal Convolutional Networks (TCN), were implemented and compared. Their performance was assessed based on sensitivity (S), positive predictive value (PPV), and training time. Among these, biLSTMs and TCNs demonstrated superior segmentation accuracy, achieving sensitivity and PPV values of approximately 90% and 80%, respectively. The neural networks developed in this study offer a universal framework for signal segmentation across varying chip architectures, addressing previous challenges of robustness and generalizability. By enabling efficient, high-speed processing, the integration of MIC and ML lays the groundwork for innovative single-cell workflows, with applications in diagnostics, drug discovery, and environmental monitoring.