Morphological Clustering of Cell Cultures Based on Size, Shape, and Texture Features

High content screening for drug discovery in cancer research relies increasingly on cell-based models, using microscopic imaging as a primary readout. In combination, microscopic imaging and cell culturing provide powerful tools for studying cancer-relevant cell biology in vitro. As a result, an enormous amount of complex biometric image data is generated that can be used for high throughput and high content analyses. We present a method for computationally efficient and flexible quantification of multicellular structures or tumor spheroids, conducted in a semi-unsupervised manner. Our phenotypic clustering approach is based on morphological features, in particular, on size and novel shape and texture features. It consists of multiple automated steps in which the information characterizing the most relevant morphological features is first extracted from the images, the dimension of the features is reduced, and finally, structures are clustered into biologically meaningful groups. Local central moments and local binary operators characterize the texture, whereas shape features are obtained by an alignment to elliptical and smooth reference shapes. Using simulation studies, we show that the cluster identification performs well and demonstrates good repeatability in the presence of random orientation, size, rescaling, and texture. We show how the method can be applied to an actual high-content imaging dataset to find an intuitive and flexible summary of high content screens, not achievable with existing tools. Supplementary materials for this article are available online.

癌症研究中用于药物研发的高内涵筛选（High Content Screening）愈发依赖基于细胞的模型，并以显微成像作为核心读出手段。显微成像与细胞培养相结合，为体外研究癌症相关细胞生物学提供了强有力的工具。由此产生了海量复杂的生物特征图像数据，可用于高通量与高内涵分析。本文提出一种采用半无监督方式实现的、计算高效且灵活的多细胞结构与肿瘤球状体（tumor spheroids）量化方法。我们的表型聚类（phenotypic clustering）方法基于形态学特征，尤其聚焦于尺寸、新型形状与纹理特征。该方法包含多步自动化流程：首先从图像中提取表征关键形态学特征的信息，随后对特征维度进行降维，最终将结构聚类为具有生物学意义的组别。局部中心矩与局部二值算子用于表征纹理特征，而形状特征则通过与椭圆型光滑参考形状的对齐操作获取。通过仿真实验，我们验证了该聚类识别方法性能优异，且在存在随机取向、尺寸变化、缩放及纹理差异的场景下仍具备良好的可重复性。我们还展示了该方法如何应用于真实的高内涵成像数据集，以实现现有工具无法达成的、对高内涵筛选结果的直观且灵活的总结。本文补充材料可在线获取。