Neuromorphic Imaging Cytometry on Human Blood Cells

Abstract

AO_SCPLOWBSTRACTC_SCPLOWImage-enhanced cytometry and sorting are powerful technologies that provide single-cell resolution and, where possible, cell actuation based on spatial and fluorescence characterisation. With the emergence of deep learning (DL), numerous cytometry-related works incorporate DL to assist their research in handling data-intensive and repetitive workloads. The rich spatial information provided by single-cell images has exceptional use with DL models to classify cells, detect rare cell events, disclose irregularity and achieve higher sample purity than a conventional feature-gating strategy. One of the significant challenges in these image-enable technologies is the constrained throughput owing to the data-expensive image acquisition and balancing between speed and resolution. This work introduces a novel paradigm by adopting a bio-inspired neuromorphic photosensor to capture fast-moving cell events. It facilitates a data-efficient, fluorescence-sensitive, fast inference approach to establish a foundation for neuromorphic-enabled cytometry/sorting applications. We have also curated the first neuromorphic-encoded cell dataset, including human blood cells (red blood cells, neutrophils, lymphocytes, thrombocytes), endothelial cells and polystyrene-based microparticles. To evaluate the data quality and potential of DL-based gating, we have directly trained a hybrid classification model based on this dataset, accomplishing a promising performance of 97% accuracy and F1 score with a significant reduction in memory usage and power consumption. Combining neuromorphic imaging and DL holds substantial potential to develop into a next-generation AI-assisted cytometry and sorting application.