Droplet Analysis – Deep Learning based

This project addresses antibiotic resistance by developing a high-throughput platform for rapid antibiotic susceptibility testing. The platform integrates microfluidic droplet encapsulation of individual bacterial cells, (ii) 2D angle-resolved light scattering (ARS) for label-free imaging, and (iii) deep learning algorithms to analyze microbial growth dynamics.

This project used advanced microfluidic and AI technology for rapid and reliable antibiotic susceptibility testing in just two hours.

Schematic of the workflow. i) Laser-based sensor for ARS imaging in a microfluidic chip. ii) ARS imaging of droplets flowing through a microfluidic channel. iii) ARS images were analyzed with a CNN. A predicted OD is given as output (whether the cells inside of the droplets grew or not).

The analysis process begins with laser-based ARS imaging of picoliter-sized droplets within a microfluidic chip. Different exposure times (800, 1500, 2000 µs) were employed for imaging droplets, and principal component analysis (PCA) was utilized to visualize distinct OD clusters. EfficientNetV2 models were specifically used for feature extraction and OD prediction to enhance the accuracy of bacterial growth detection.

Exposure time analysis. a) Droplets with high cell concentrations imaged with exposure times of 800, 1500 and 2000 µs, and the analysis steps. b) Unsupervised deep learning followed by PCA of ARS images taken with exposure time of 800 µs and 1500 µs show that clusters can be clearly separated.

The ARS images are subsequently analyzed using convolutional neural networks (CNNs) to predict optical density (OD), thereby determining bacterial growth. This approach facilitates rapid and reliable antibiotic susceptibility testing, supporting high-throughput screening and offering improved sensitivity compared to traditional method.

Overview of the analysis steps of ARS images. Predicted OD distributions versus mean ODs of 5 different droplet populations with S. aureus cells.
Violin plot of predicted OD at each timepoint of ARS imaging and Brightfield imaging. Significant growth was visible in ARS images after only 1-2 hours, whereas for Brightfield imaging, significant change was only visible after 3 hours.

The platform employs microfluidic droplet encapsulation, angle-resolved light scattering (ARS), and deep learning algorithms to detect bacterial growth and antibiotic susceptibility in a time frame of one to two hours, a significant improvement over our previous methods. We demonstrate the efficacy of the method indicating the potential for transformative improvements in clinical diagnostics.

Experimental Collaborators

HKI Bio Pilot Plant at the Leibniz-HKI in Jena, Germany.

Publications

Posters

12335 Next