R. Krampl, A. Trcka
Keywords: AI, camera, microscope, cells, bioreactor, lab, cuvette, automation, auto-labeling
Summary:Recent advancements, standardization and process definition of image recognition have enabled the automation of highly repetitive tasks in multiple fields. Characterization and auto-labeling of live image streams is now possible with a relatively small development team using commercial off-the shelf computational electronics, coupled with application specific HW such as cameras and sensors. In-situ monitoring of bioreactor products of all sizes is now possible without the need to take samples, which are manually demanding and have the potential to introduce process errors or contamination. In this study we present an add-on system to generic bioreactors which utilizes machine learning tools to continuously sample and detect anomalies within a bioreactor stream. This is achieved by utilizing a glass cuvette that is paired with a commercial camera mounted on a microscope in its most basic format, covering the visible spectrum. The system allows for active image cognition of any biological cells like bacteria, yeast, algae in a continuous mode without any sample extraction by fully recycling the fluid back to the original bioreactor. The ML model is processed (parameterized) on a PC workstation, and once tuned it is deployed on a small footprint, remote Embedded AI Computer, assuming that operational needs require larger production facilities, as opposed to deployment in a highly controlled lab.