A deep learning-based concept for high throughput image flow cytometry

Abstract

We propose a flow cytometry concept that combines a spatial optical modulation scheme and deep learning for lensless cell imaging. Inspired by auto-encoder techniques, an artificial neural network mimics the optical transfer function of a particular microscope and camera for certain types of cells once trained and reconstructs microscope images from simple waveforms that are generated by cells in microfluidic flow. This eventually enables the label-free detection of cells at high throughput while simultaneously providing their corresponding brightfield images. The present work focuses on the computational proof of concept of this method by mimicking the waveforms. Our suggested approach would require a minimum set of optical components such as a collimated light source, a slit mask, and a light sensor and could be easily integrated into a ruggedized lab-on-chip device. The method is benchmarked with a well-investigated dataset of red blood cell images.

Effect of Cell Age and Membrane Rigidity on Red Blood Cell Shape in Capillary Flow

“…Deviations from the equilibrium RBC distributions have previously been used to assess changes in RBC flow behavior in patients with neuroacanthocytosis syndrome and COVID-19, as well as in patients before and after undergoing hemodiafiltration dialysis [34,35]. Moreover, RBC shapes in microfluidic capillary flow have been studied for healthy and diseased RBCs using manual and machine learning shape classification approaches [12,[36][37][38][39][40]. Recent studies have shown the potential of microfluidic characterizations of RBC shapes as biomarkers for specific pathologies, to assess the cell deformability [34,41,42].…”