Machine learning-based bead enumeration in microfluidics droplets enhances the reliability of monitoring bead encapsulation toward single-cell sorting applications

The encapsulation of cells within droplets is a crucial aspect of various cell analysis applications. Current research has focused on accurately detecting and identifying cell types or cell counts within droplets using object detection in bright-field images. However, there are only a few in-depth investigations into the impact of the image data quality acquired from optical systems on computer vision models. This study examines several popular machine learning object detection models to analyze scenarios complicating the identification of bead locations within a droplet, posing challenges for computer vision models. A microfluidic droplet generation system was developed and implemented, coupled with optical devices to capture images of encapsulated beads within the droplet. To identify the most efficient model, a specific dataset was meticulously selected from the overall data, encompassing images depicting overlapping beads and edge-drifting scenarios. The proposed method achieved up to 98.2% accuracy on the testing dataset and 95% in real-time testing with the YOLOv8 model, enhancing bead count precision within droplets and clarifying the correlation between accuracy and frame recognition thresholds. This work holds particular importance in single-cell sorting, where precision is critical in ensuring meaningful outcomes, particularly concerning rare cell types such as cancer cells.