A High Performance of Single Cell Imaging Detection with Deep Learning

Single cell imaging enables new applications such as biomedical diagnostics, food inspection, and water quality monitoring. In this paper, we study the cell imaging-based machine learning techniques for high-performance cell detection. By taking the advantage of deep learning and imaging flow cytometry, we manage to detect Cryptosporidium and Giardia cells in the bright-field images with high accuracy and high speed on embedded GPU system. Our experiments demonstrate that the newly developed deep learning-based algorithms surpasses the hand-crafted features and SVM-based algorithms. We achieved above 99 percentage in accuracy and 580+fps in speed on embedded Jetson TX2 platform. Our research will lead to a highly accurate real-time single cell level detection system in future.

I. Introduction

Single cell imaging is an important technique for applications in diverse fields including disease diagnostics as well as food and water safety and security. In nature, cells at different stages of life have distinct size, morphology as well as density. The morphology of healthy protozoan and bacterial cells is distinctively different from those in apoptotic or necrotic states. Furthermore, images of microbes in water biome exhibit distinctive features. For example, Cryptosporidium have a spherical shape whereas Giardia appear round to ovoid. Therefore, by scrutinizing images of single cells, important information on their species and stage of life cycle is revealed [1]. Especially in water industry, this single-cell imaging technique plays an important role in water quality, safety and security which are highly influenced by the quality control of drinking water. According to USEPA standard, 10 L of drinking water needs to be tested on daily basis for potential protozoa contamination [2]. Immunolabeling and microscopic imaging are the golden standard for protozoa analysis according to USEPA 1623.1. This approach takes at least 1.5 days and results in a low accuracy due to the hands-on protocol. Furthermore, there are millions of species in aquatic microbiota with different sizes ranging of sub-micron to millimeter. Compared to the immunolabeling-based technique which is limited to the detection of known species, label-free imaging offers a holistic technique for the identification of all species, as well as for the analysis of cell size, shape and density.