Automated droplet classification using computer vision

I developed multiple high-throughput computer vision algorithms for detecting and classifying ice nucleation or freezing of picolitre droplets in microfluidic channels. Initial trials with feature-based machine learning algorithms proved not sufficiently accurate. I addressed this challenge by developing a deep neural network (DNN) based image recognition approach to achieve accuracy >99.5% compared to human classification. I integrated this DNN into a custom MATLAB GUI app for usability and wider adoption by other researchers. I used this network to collect millions of data points and reduce analysis time 100-fold resulting in about 200k droplets per data point.

I developed another approach to handle high-speed (10,000+ fps) video for a real-time sorting project, as the DNN cannot process high-speed camera output at runtime. I solved this problem by designing a polarizer-analyzer passive optical system to detect frozen droplets to complement the more accurate DNN.

This classifier is currently being used to develop a droplet sorting system with the assistance of a active flow control system.

Publications

• Roy P., House M. & Dutcher C.; Multi-pronged approach for high throughput detection of ice nucleation in a microfluidic device. Micromachines, In preparation.

• Roy P. & Dutcher C.; Approaches for High Throughput Ice Nucleation Detection in a Microfluidic Platform. 85th New England Complex Fluids Meeting. Virtual workshop. 3 minute presentation.

• Roy P. & Dutcher C.; Approaches for high throughput ice nucleation in a microfluidic platform. Annual Meeting of Center for Aerosol in the Chemistry of the Environment (CAICE), UC San Diego, November 4-6, 2020, Virtual Conference. Virtual presentation.