Dynamic video recognition for cell-encapsulating microfluidic droplets.

Area of Science:

• Biomedical Engineering
• Microfluidics
• Machine Learning

Background:

• Droplet microfluidics is crucial for high-throughput biomedical applications like single-cell sequencing.
• Accurate droplet size and cell encapsulation are vital for reliable results but challenging to control.
• Current machine learning methods require extensive pixel-level annotation for training.

Purpose of the Study:

• To develop and validate a weakly supervised cell-counting application (WSCApp) for microdroplet video analysis.
• To enable real-time quality control of droplet microfluidics by identifying droplet and cell locations.
• To reduce the annotation burden for machine learning model training in this field.

Main Methods:

• Implemented a weakly supervised cell-counting network (WSCApp) for video recognition of microfluidic droplets.
• Applied the software to process videos of microfluidic droplets encapsulating various cell types and beads.
• Utilized transfer learning to fine-tune a pre-trained model, minimizing annotation requirements.

Main Results:

• WSCApp demonstrated real-time video processing capabilities for microfluidic droplets.
• The software accurately identified droplet locations and encapsulated cells without supervised location data.
• Achieved high accuracy in distinguishing droplet encapsulations (micro-F1 score > 0.94).
• Transfer learning reduced annotation effort by over 80%.

Conclusions:

• WSCApp offers an effective solution for automated quality control in droplet microfluidics.
• The software facilitates accurate cell counting and location identification in microdroplets.
• This approach significantly lowers the barrier to using machine learning for microfluidic applications.