Deep learning methods for the analysis of microfluidic and acoustofluidic systems

In this work, we developed a series of innovative Deep Learning (DL) based approaches designed to enhance or enable the characterization of microfluidic and acoustofluidic phenomena and devices. General Defocusing Particle Tracking (GDPT) is one of the primary velocimetry methods used in this field for determining the three-dimensional (3D) trajectories of tracer particles. GDPT relies on a single-camera defocusing principle and requires monodisperse tracer particles, but it is unsuitable for cases where particles are non-spherical or vary significantly in size. To overcome these limitations, we developed a hybrid method that combines defocusing with DL. We adapted the VGG-16 (16-layer Visual Geometry Group) architecture for a regression task and demonstrated that the proposed model can accurately estimate both depth position and orientation of non-spherical particles. We then applied the method to the study of microswimmer kinematics. Despite the complexity, the system successfully predicted its three-dimensional position and orientation. Next, we integrated our model with YOLO (You Only Look Once), a real-time object detection algorithm, for in-plane (XY) coordinate estimation and applied it to spray droplet measurements. In this case, the particles were spherical droplets with varying diameters, and the approach successfully extracted their 3D positions and sizes. We further adapted this technique for thin oil-film interferometry to estimate wall shear stress. Results showed that the method could detect droplets and determine the average fringe distance in real-time. Finally, we used unsupervised machine learning to classify different stages of 3D cell aggregation under acoustic standing waves. ResNet-50 was utilized for feature extraction, followed by K-Means clustering to analyze high-resolution microscopy images and determine the time point at which a spheroid becomes fully developed. Overall, these studies demonstrate the great flexibility of DL-based algorithms in effectively addressing complex challenges in micro-acoustofluidic research, providing new, powerful tools for future advancements in the field.