Fast label-free recognition of NRBCs by deep-learning visual object detection and single-cell Raman spectroscopy

Abstract

Nucleated red blood cells (NRBCs) as a type of rare cell present in an adult's peripheral blood is a concern in hematology, intensive care medicine and prenatal diagnostics. However, it is labor-intensive to screen such rare cells from real complex cell mixtures especially in a label-free way. Herein, we report a new label-free method that incorporates image recognition and Raman spectroscopy for fast recognition of the rare cells in blood. First, we identified unlabeled NRBCs based on both Raman signals of hemoglobin and nucleated morphology, and recorded their microscopic image characteristics which were different enough from other blood cells in unlabeled morphology. Then, two deep-learning algorithms of visual object detection, Faster RCNN and YOLOv3, were investigated for cell morphological recognition on a low-cost computer configuration, and YOLOv3 was demonstrated to be more competent for real-time detection despite slightly lower precision. Finally, several NRBCs were successfully found in maternal blood using this method, which verified the methodological feasibility. Thus, we believe such a labor-saving approach might inspire a new idea for detecting rare cells from complex cell mixtures in a label-free and computer-assisted way.