AN ARTIFICIAL INTELLIGENCE FOR RAPID IN-LINE LABEL-FREE HUMAN PLURIPOTENT STEM CELL COUNTING AND QUALITY ASSESSMENT

Abstract

The current state-of-the-art in hPSC culture is a bespoke and user-dependent process limiting the scale and complexity of the experiments performed and introducing operator-to-operator and day-to-day variation. Artificial intelligence (AI) offers the speed and flexibility to bridge the gap between a human-dependent process and industrial-scale automation. We evaluated an AI approach for counting exact cell numbers of undifferentiated human induced pluripotent stem cells in brightfield images for automating hPSC culture. The neural network generates a topological density map for accurate cell counts. We found that the image-based AI algorithm can determine a precise number of hPSCs and is superior to fluorescence-labeled object detection; the algorithm can ignore well edges, meniscus effects, and dust, achieving an average error of 5.6%. We have built a prototype capable of making a go/no go decision for stem cell passaging to perform 26,400 individual well-level counts from 422,400 images in 12 hours at low cost.