Image-based analysis of cells is a powerful modality to measure and record the high content information which reflect the cellular status during their culture. By the recent advances of image processing and machine learning technology lead by artificial intelligence (AI) applications in other industrial fields, “image data” will be a new frontier of
biological big data which will support the cell-related life sciences. However, cell images around the world have not been standardized among all the cell-handing facilities, since
microscopic observation skills and techniques have been the most trusted method to control cell culture process for more than 100 years. Such situation is greatly different from the other industrialized fields, which challenged to abandon the experts’ feeling-based manufacturing manner to evolve to the data-driven manufacturing. To introduce the most advanced technologies for mechanization and automation of cell culture processes to maximize the efficiency and stability of cell-related techniques, we have to prepare for the next stage after the simple introduction of novel technologies in cell culture, the period of “data integration for data-driven activities”. Our group has been reporting the label-free
morphology-based analysis approaches for developing enabling technology for cell quality monitoring and control for maximizing the efficiency and reproducibility of cellular
researches and manufacturing [1-3]. However, there are still rare researches that report the possibilities and effectiveness of studying the “data integration” of collected “cell images”.
Therefore, in this work, we developed the data integration technology for enhancing the future coming “image-based cell quality control process” in cell manufacturing industry.
[Methods] We examined the effect of data bias effect between image differences and facility differences within the same culture process for manufacturing mesenchymal stem
cells for therapeutic use. Especially, we examined the image resolution effects, and developed effective image interpolation method for standardizing the orphological
descriptor effects for predicting cell yield of mesenchymal stem cells.
[Results and conclusions]
From our data, we found that the bias of lighting effect is one of the most crucial noises that disturbs the morphology-based machine learning model performances. Moreover, we also found that the resolution differences between cell culturing facilities can be interpolated effectively to achieve equivalent cell yield prediction performances. This investigation will suggest the importance of our present situation which
neglect the image data quality in the data accumulation process for further AI applications. Our data also indicates that morphology-based cell image analysis has higher potency of obtaining robust prediction models compared to the deep network models, since it enables detailed data normalization.
- Sasaki, H., et al., PLoS One 9(4), e93952 (2014)
- Kato R., et al., Sci. Rep., 6, 34009, (2016)
- Imai Y., et al., Inflamm. Regen. 42(1), 8-20 (2022)