Convolutional neural networks-based red blood cell detection and tracking with automatic data generation

Abstract

Understanding the flow characteristics of red blood cell (RBC) is crucial to comprehending the oxygen supply mechanism of microcirculation. Quantification of the RBC partitioning in bifurcating channels is necessary; however, most measurements are often performed manually, which has limitations in terms of labor intensive and reproducibility. Existing automatic detection methods are insufficient to identify RBCs obtained from in vitro experiments, due to heterogeneous backgrounds, minimal luminance variations, and unclear contours of RBCs. Furthermore, because RBCs are deformable, a method capable of tracking numerous moving and deforming RBCs is also required to investigate how capillary networks influence RBC deformation. We developed a convolutional neural networks (CNNs)-based method for detecting and simultaneously tracking multiple deformable RBCs in images obtained from in vitro experiment. The target images were obtained from a microfluidic channel with a ladder structure to understand the RBC heterogeneity mechanism in capillary networks. We also developed a method for automatically generating pseudo-RBC images based on actual RBC images to train CNNs. Moreover, we used the difference image between consecutive frames, along with RBC images, as inputs for CNNs. We validated the detection accuracy and evaluated the hematocrit and flow rate results obtained using the proposed method for each channel during in vitro experiments. The proposed method exhibited exceptional performance, with a precision of ≥0.925, F₁ score of ≥0.825, and low false positive rate across all frames of the RBC images used for testing. The proposed detection method enables the automated, accurate, and high-throughput quantification of RBC positions within in vitro experiments, facilitating more quantitative assessments of RBC flow characteristics within capillary networks. Moreover, it can be readily adapted for automated tracking of various cell types beyond RBCs, as well as in vivo experiments. This may lead to a more detailed understanding of microcirculatory dynamics and related physiological processes.

GraphicalAbstract