Deep learning supported image analysis of angle-resolved scattered light images of bacteria in microfluidic droplets

www.leibniz-hki.de References Deep learning supported image analysis of angle-resolved scattered light images of bacteria in microfluidic droplets Arjun Sarkar 1,2, Carl-Magnus Svensson 1, Martina Graf 2,3, Anne-Sophie Munser 4, Miriam A. Rosenbaum 2,3 , Marc Thilo Figge 1,2 1Applied Systems Biology, Leibniz Institute for Natural Product Research and Infection Biology –Hans Knöll Institute, Jena, Germany 2Friedrich Schiller University, Jena, Germany 3 Bio Pilot Plant, Leibniz Institute for Natural Product Research and Infection Biology –Hans Knöll Institute, Jena, Germany 4Fraunhofer Institute for Applied Optics and Precision Engineering IOF, Jena, Germany [1] Schröder S et al. 2011. Applied Optics. 50(9):C164-C171 [2] Tan, M., & Le, Q. 2021. ArXiv, abs/2104.00298 Microfluidic droplets and angle-resolved scatter (ARS) images Training DL regression model and Growth analysis Growth analysis after antibiotic (TET) application •Angle-resolved scattered light imaging (ARS) gives fast and highly resolved information about structures and objects [1] •ARS applied to picoliter-sized droplets in flow to detect cell growth on a single-cell level •Two bacteria species - Staphylococcus aureus and Escherichia coli •Traditional image analysis and Deep Learning (DL) technique to quantify changes in spectra •Goal is to detect cell division events for rapid antibiotic susceptibility testing [email protected] Work is funded by BMBF through InfectoGnostics 2 (ADA Nr. 13GW0456B) Exposure-Time Prediction Falsely-triggered Image Removal Growth Analysis Classification of 800 µs, 1500 µs, 2000 µs exposure images –to determine the fastest exposure images that can be classified using DL. DL network was trained to classify and remove falselytriggered images. DL regression network was trained on images to predict the optical density (OD). Trained model was applied to the growth dataset (0h to 5h). Workflow Data analysis of droplets - control and different TET concentrations Empty droplet Droplet with S. aureus Droplet with E.coli ARS images of droplet with S. aureus and E. coli Results - regression on S. aureus on the test set DL model - falsely triggered image classifier CNN Regression DL model –trained on dilution series (ODs) Falsely triggered images removed from dataset Supervised DL based regression ARS images of droplets at different timepoints for growth analysis 0 h 1 h 2 h 3 h 4 h 5 h Growth Analysis after antibiotic application Application of TET (Tetracycline) –antibiotic (0.6, 2, 10 µg/mL). Growth was analysed after TET application. Growth analysis across different timepoints (S. aureus –threshold > 0.1 OD) Selected 800 µs and 1500 µs images for model training Confusion matrix Supervised DL based regression •Trained falsely-triggered classifier removes all falsely-triggered images from dataset •DL regression model (EfficientNetV2) [2] trained on OD values (OD 0, OD 0.01, OD 0.1, OD 0.5 and OD 1) Growth analysis •Dataset with ARS images from hour 0 to hour 5 •Trained DL regression model predicts OD of each droplet in the dataset •Threshold set at OD 0.1 as model fails to correctly predict empty/non-empty droplets below OD 0.1 Comparison between brightfield image and ARS image •In brightfield imaging, slightly significant growth visible after 3 hours •In ARS imaging, with DL analysis, significant growth visible from 2 hours TET application and growth analysis •Dataset with 3 TET concentrations –0.6, 2, and, 10 µg/mL •Trained DL regression model predicts growth (OD) •No significant growth reduction after application of TET concentration - 0.6 µg/mL (7% of the total droplets show growth) •Significant reduction in growth from TET concentration of 2 µg/ML - 0.5% of the total droplets show growth •At TET concentration 10 µg/ML only 0.04% of the total droplets show growth