Single-frame multi-stage transformer-based segmentation network for droplet localization and volume prediction in digital microfluidics

Abstract

Digital microfluidics (DMF) enables precise control of droplets for tasks like transportation, mixing, and analysis. However, current DMF systems rely on manual control, which is time-consuming and inefficient. Existing visual-based automated control systems typically depend on simple location techniques, which require stable lighting, multiple frames, or reference images, limiting their practicality and robustness. This paper proposes a single-frame multi-stage Transformer-based segmentation network (SMTSN) for multi-target segmentation of droplets and electrodes. Despite a low contrast (0.0177 Weber Contrast) between droplets and surrounding silicone oil, SMTSN can still achieves an $A{P}_{50:5:95}$ of 0.875 for droplet segmentation and 100 % accuracy in droplet localization. The network is integrated with the DeepSort tracking algorithm, enabling automated control of droplet operations such as merging, agitation, and separation. We conducted automated detection of five kinds of viral diarrhea in children across 1005 cases. Our results show 100 % consistency with the results of genetic sequencing.