With rapid advances in cell therapy, technologies enabling both consistency and efficiency in cell manufacturing are becoming necessary. Morphological monitoring allows practical quality maintenance in cell manufacturing facilities, but relies heavily on human skill. For more reproducible and data-driven quality evaluation, image-based morphological analysis provides multiple advantages over manual observation. Our group has investigated the performance of multiple morphological parameters obtained from time-course images to non-invasively and quantitatively predict cellular quality using machine learning algorithms. Although such morphology-based computational models succeeded in early cell quality predictions, it was difficult to introduce our approach in cell manufacturing facilities owing to data variation issues. Since manufacturing facilities have fixed their protocol to minimize anomalies as much as possible, most accumulated data are normal, and anomalies are scarce. Thus, our morphological analysis had to adapt to such practical situation where it was difficult to observe a wide range of data variations, including both normal samples and anomalies, which is typically essential to improve most machine learning models' performance. In the present study, we introduce a practical morphological analysis concept by investigating the performance of anomalous quality decay discrimination during the continuous passaging of human mesenchymal stem cells (hMSCs). Combining the visualization method and asymmetric statistic discrimination, we describe an effective morphology-based, in-process quality monitoring concept to detect quality anomalies throughout cell culture process. Our results showed that the use of morphological parameters to reflect cellular population heterogeneity can predict hMSC quality decay within 6 h after seeding.

随着细胞疗法的快速发展，在细胞制造中同时实现一致性和效率的技术变得必要。形态学监测可以在细胞生产设施中保持实用的质量，但是严重依赖于人类的技能。对于更可重现和数据驱动的质量评估，基于图像的形态分析比手动观察具有更多优势。我们的小组研究了从时程图像获得的多种形态学参数的性能，以使用机器学习算法以无创和定量方式预测细胞质量。尽管这种基于形态学的计算模型成功地进行了早期电池质量预测，但是由于数据变化问题，很难将我们的方法引入电池制造设施。由于制造工厂已固定其协议以尽可能减少异常，因此大多数累积数据都是正常的，并且异常很少。因此，我们的形态分析必须适应这样的实际情况，即难以观察到各种数据变化，包括正常样本和异常，这对于提高大多数机器学习模型的性能通常是必不可少的。在本研究中，我们通过研究人类间充质干细胞（hMSCs）连续传代过程中异常质量衰变判别的性能，介绍了一种实用的形态分析概念。结合可视化方法和非对称统计量判别，我们描述了一种有效的基于形态学的方法，过程质量监控概念可检测整个细胞培养过程中的质量异常。我们的结果表明，使用形态学参数反映细胞群体异质性可以预测播种后6小时内hMSC的质量下降。