Using machine learning to measure mortality and reproduction of springtails (Folsomia candida) in response to microplastics and additives

The effects of micro- and nanoplastic particles on terrestrial ecosystems are still understudied and only few studies on their effects on the model organism Folsomia candida exist to date. Mortality and reproduction tests with F. candida usually take a long time and require extensive manual counting of offspring individuals. To optimize reproduction bioassays, we trained a Fast Region-based Convolutional Neural Network (Fast-RCNN) on the task to automatically count both adults (endpoint mortality) and offspring (endpoint reproduction) from images taken at the end of experiments. By comparing the results from automatic measurements with manual counts, we demonstrate the accuracy and precision of the model and show how the model can be used to make the measurement process faster and less work intense. We finally apply the trained model to outcomes of a feeding experiment testing for effects of seven different microplastic particles manufactured and characterized by the Collaborative Research Center (CRC) 1357 Microplastics at the University of Bayreuth: polystyrene (PS) particles without added additives, PS particles containing controlled amounts of the additive Irgafos 168 (low and high concentration), PS particles containing controlled amounts of the additive Irganox 1010 (low and high concentration) and PS particles with increased contents of styrene monomers. Images were taken after four weeks of exposure and analyzed automatically with the trained model. The results of the experiment show a consistent but small increase in mortality in springtails exposed to all tested PS particles, but no effect on offspring numbers per surviving individual. The trained model will be made openly available upon publication. Also see: https://micro2024.sciencesconf.org/559677/document