A hybrid neural network and sparse recovery framework for resolving multi-cell coincidence in label-free impedance cytometry.

Microfluidic impedance cytometry (MIC) is a critical technology for high-throughput single-cell analysis owing to its label-free, non-invasive operation and ease of integration. However, under high-throughput conditions, baseline drift, noise, and multi-cell coincidence events distort cell counting and amplitude estimation. To enhance detection accuracy, this work presents a single-cell signal detection system based on a coplanar three-electrode differential impedance microchip, in which deep learning is integrated with a sparse recovery algorithm. Raw impedance signals acquired by the three-electrode differential configuration are first processed by a recurrent neural network (RNN) to automatically identify and precisely segment cell events. Subsequently, a one dimensional convolutional neural network (1D-CNN) is introduced to classify the number of coincident particles in each event segment, and the classification results are used as the sparsity prior for the sparse recovery algorithm. Based on this prior, an -norm-based sparse recovery algorithm separates multi particle coincident waveforms and accurately recovers single-cell amplitudes. Validation experiments using human lymphoma cells verify that our system achieves precise particle event segmentation, reliable identification of coincident particle counts, and high-precision recovery of single-particle amplitude characteristics. Overall, the proposed detection framework effectively improves the accuracy and operability of MIC single-cell analysis in the presence of coincident particles, laying a scalable technical foundation for high-throughput and intelligent single-cell measurement in subsequent studies.