A machine vision perspective on droplet-based microfluidics

Abstract

Microfluidic droplets, with their unique properties and broad applications, are essential in in chemical, biological, and materials synthesis research. Despite the flourishing studies on artificial intelligence-accelerated microfluidics, most research efforts have focused on the upstream design phase of microfluidic systems. Generating user-desired microfluidic droplets still remains laborious, inefficient, and time-consuming. To address the long-standing challenges associated with the accurate and efficient identification, sorting, and analysis of the morphology and generation rate of single and double emulsion droplets, a novel machine vision approach utilizing the deformable detection transformer (DETR) algorithm is proposed. This method enables rapid and precise detection (detection relative error < 4% and precision > 94%) across various scales and scenarios, including real-world and simulated environments. Microfluidic droplets identification and analysis (MDIA), a web-based tool powered by Deformable DETR, which supports transfer learning to enhance accuracy in specific user scenarios is developed. MDIA characterizes droplets by diameter, number, frequency, and other parameters. As more training data are added by other users, MDIA's capability and universality expand, contributing to a comprehensive database for droplet microfluidics. The work highlights the potential of artificial intelligence in advancing microfluidic droplet regulation, fabrication, label-free sorting, and analysis, accelerating biochemical sciences and materials synthesis engineering.