Deep learning powered biophysical image-based flow cytometer
Mechanical properties of single cells are widely associated to their biological activities and dysfunction which can be exploited for clinical diagnosis. While microfluidic deformability cytometers are emerging tools for high throughput single cell measurements, image analysis is often laborious whi...
Saved in:
Main Author: | |
---|---|
Other Authors: | |
Format: | Final Year Project |
Language: | English |
Published: |
Nanyang Technological University
2022
|
Subjects: | |
Online Access: | https://hdl.handle.net/10356/157650 |
Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
Institution: | Nanyang Technological University |
Language: | English |
Summary: | Mechanical properties of single cells are widely associated to their biological activities and dysfunction which can be exploited for clinical diagnosis. While microfluidic deformability cytometers are emerging tools for high throughput single cell measurements, image analysis is often laborious which hinders their clinical applications. In this thesis, we report a novel deep learning powered microfluidic deformability cytometer for high-throughput cellular mechanophenotyping. The microfluidic device consists of a cross-junction whereby the extensional flow of viscoelastic medium deforms the cells to enable image quantification of dynamic cell morphology and deformability. Machine learning through the application of convolutional neural networks (CNNs) will be deployed to develop a classification model of blood cells with different cell phenotypes. We performed three experiments, namely: HL60 – untreated and treated with Paraformaldehyde (PFA); neutrophils – untreated and treated with Phorbol Myristate Acetate (PMA); and red blood cells (RBCs) and white blood cells (WBCs). Cell images were extracted using OpenCV in Python and analyzed using both a Convolutional Neural Network (CNN) and the You Only Look Once (YOLO) v4 model. Overall, our results showed CNN model accuracies and YOLOv4 confidence levels of at least 80% to distinguish different cell phenotypes based on morphological and deformability changes at various sheath flow rates. Further work includes the integration with microfluidic modules to offer machine learning-enabled real-time cell analysis for point-of-care diagnostics. |
---|