Automatic Sorting of Zebrafish Embryos using Deep Learning

Many of biological studies like transcripromics or metabolomics requires a large number of zebrafish embryos. Dead or unfertilized embryos that will not be useful for studies should be eliminated. Biologists frequently perform this manually, which is laborious, error-prone, and time consuming. We therefore proposed a method for sorting these undesired cells using deep learning and microfluidics. A YOLOv5 model was trained with a 95% accuracy and a processing speed of 10.6 ms per frame to assess the stage of development as well as whether a zebrafish egg is dead, unfertilized, or alive. Eggs are housed in traps on a microfluidic chip using micro-pumps. Once all the zebrafish eggs are housed in the traps, the microfluidic chip is placed in an XYZ motorized stage which, by moving, allows the detection of the eggs by the deep learning system and automatically sorting them based on dead or unfertilized embryo detected. The sorting experiment was conducted in two modes: without feedback and with feedback while using the dead egg position. The first one had a sorting success rate of 90% as opposed to 100% for the feedback mode with 3 seconds required for each dead egg.