CellAi® SOFTWARE ENABLES LABEL FREE DETERMINATION OF VIABILITY FROM BRIGHTFIELD IMAGES OF MSC USING ARTIFICIAL INTELLIGENCE NEURAL NETWORKS

• Published in: Cytotherapy (Oxford, England) Vol. 26; no. 6; p. S141
• Main Authors: Piemontese, M., Krasniqi, F., Aldeghi, G., Davies, S., Dobson, P.
• Format: Journal Article
• Language: English
• Published: Elsevier Inc 01.06.2024
• Subjects: Advanced Basic Science; artificial intelligence; label-free stain; Other; viability; artificial intelligence; label-free stain; viability
• ISSN: 1465-3249; 1477-2566
• DOI: 10.1016/j.jcyt.2024.03.275

Cell number and viability measurements provide critical information for decision-making throughout development and manufacturing of cell-based products. Gold standard methods to measure count and viability of cells include manual processes such as Trypan Blue methods, or semi- automated platforms leveraging fluorescent stains. These processes can be slow, low throughput, and are subject to operator error. We aim to displace current approaches with our CellAi® software that enables label free determination of live/dead cells from brightfield images using Artificial Intelligence neural networks. To create a training dataset of labelled live and dead cells, MSC were stained with DAPI and CalceinAM, seeded in 96 well plates and incubated for 10 minutes. MSC were imaged using the SparkCyto imager (Tecan) at 10X in the brightfield, blue and green channel. In parallel, cell number and viability were recorded from the same sample using the NucleoCounter® NC-200 (ChemoMetec) and manual Trypan blue method, to benchmark model performance. Finally, we identified a set of variables relevant to the MSC workflow (i.e., presence/absence of CPA, viability range, cell density) and generated specific datasets to ensure model generalization. Following segmentation of single cells using CellPose, we assigned pseudo-labels to each cell based on CalceinAM and DAPI staining. Computed pseudo-labels were used to train a deep neural network to detect live/dead cells from the brightfield images alone. Predicted viability values were benchmarked against NucleoCounter measurements, and a mean difference in viability of less than 5% was observed in test sets containing images of MSCs with both low and high viabilities (15-90%). In addition, extraction of morphological features identified specific features (e.g. standard deviation of intensity) that can be used to differentiate live and dead cells, and to validate CellAi® model predictions. The CellAi® software solution allows cell analysis with minimal sample manipulation, providing the opportunity to multiplex with additional analysis and deliver shorter, simpler lab protocols. This approach further enables integration into automated workflows, which we aim to demonstrate on the AutoCRAT platform, designed for iPSC culture, iMSC differentiation and cell characterization. Morphological feature extraction allows for the identification of critical cell features that can be exploited for deeper cell characterization during cell processing.