Segmentation of single-cell impedance signals using deep learning: a multi-dataset study

Microfluidic impedance cytometry (MIC) is a high-throughput, label-free technology for single-cell analysis, with applications in cell classification and non-invasive monitoring. However, developing a universal signal-processing approach for MIC is challenging, because signal characteristics strongly depend on the experimental setup. This study investigates the integration of MIC with deep learning (DL) to enable effective processing of raw impedance signals from various MIC systems. Specifically, the research focuses on signal segmentation (i.e., event detection), which represents the first step of the processing pipeline and therefore influences the entire workflow.

Impedance data from multiple experimental setups were collected. They are characterized by raw traces with diverse attributes and event signals exhibiting distinct temporal shapes, thus forming a rich and comprehensive database. Several DL models were implemented and compared, including recurrent, convolutional, and encoder-decoder neural networks.

While all models demonstrated good segmentation performance, the encoder-decoder network outperformed the others, achieving a sensitivity and positive predictive value of 91.6% and 91.8%, respectively. Moreover, the network remained robust when, after training, validation, and testing, it was further evaluated on additional previously unseen data.

We developed a framework for signal segmentation in MIC that addresses the challenge of cross-setup generalizability.

By enabling efficient, high-speed processing, the integration of MIC and DL lays the foundation for next-generation single-cell workflows with applications in diagnostics, drug discovery, and environmental monitoring.