The ratio between living and dead cells is an important parameter in microalgae culture and environmental monitoring. Fast, robust and automated analytical methods for monitoring microalgae growth for biotechnological and pharmaceutical applications and for optimizing production strains are needed. We developed a pipeline for the automatic binary classification of living and dead microalgae. We compared multispectral fluorescence microscopy and flow cytometry as readout platforms. Images of Chlorella vulgaris suspension cultures were captured and features of microalgaes (e.g., size) were extracted by bioimage informatics. We classified the microalgae cells with seven machine learning algorithms, including Naive Bayes, Random Forest and Neural Networks. Random Forest was a particularly suitable method for classifying the living/dead microalgae population, the label-free classifier can reach 86% of the area under the curve (AUC) if only morphological characteristics are used. The AUC rose to 99.6% if the fluorescence signal of Syto 9 (dye for all cells) and Propidium iodide (dye for damaged cells) were added as features, and the prediction accuracy reached 96.6% which is over the accuracy of 95% from the cell viability analysis using the staining method. This means that our classifier is useful as it improves the staining method by correcting 30% of the false positive/negative cells. Instead of only predicting single cells, we developed a Random Forest model to classify the distribution of living and dead microalgae populations using the fingerprint vector features, which gives an overall AUC of 94.5% and accuracy of 82%. We found that dead cells are significantly larger in average diameter (1.4-fold, p < 0.01) and area (1.5-fold, p < 0.01) when compared to living cells. In summary, combinations of statistical methods, bioimage informatics and machine learning are useful approaches for automated investigation in analyzing population dynamics of microalgae cultures.

活细胞与死细胞之间的比例是微藻培养和环境监测中的重要参数。需要用于生物技术和制药应用以及优化生产菌株的监测微藻生长的快速，可靠和自动化的分析方法。我们开发了用于对活和死微藻进行自动二进制分类的管道。我们比较了多光谱荧光显微镜和流式细胞仪作为读出平台。小球藻的图像捕获悬浮培养物，并通过生物图像信息学提取微藻的特征（例如大小）。我们使用七种机器学习算法对微藻细胞进行了分类，包括朴素贝叶斯，随机森林和神经网络。随机森林是一种用于对活/死微藻种群进行分类的特别合适的方法，如果仅使用形态特征，则无标记分类器可以达到曲线下面积（AUC）的86％。如果添加Syto 9（所有细胞的染料）和碘化丙啶（受损细胞的染料）的荧光信号作为特征，则AUC升至99.6％，预测准确度达到96.6％，超过95％的准确度。使用染色法进行细胞活力分析。这意味着我们的分类器非常有用，因为它可以通过校正30％的假阳性/阴性细胞来改善染色方法。我们不仅开发了单个森林，还开发了一个随机森林模型，使用指纹矢量功能对生活和死亡微藻种群的分布进行分类，其总体AUC为94.5％，准确度为82％。我们发现，与活细胞相比，死细胞的平均直径（1.4倍，p <0.01）和面积（1.5倍，p <0.01）明显更大。总而言之，统计方法，生物图像信息学和机器学习的组合是用于分析微藻培养种群动态的自动化调查的有用方法。我们开发了一种随机森林模型，使用指纹矢量特征对活和死微藻种群的分布进行分类，该模型的总体AUC为94.5％，准确度为82％。我们发现，与活细胞相比，死细胞的平均直径（1.4倍，p <0.01）和面积（1.5倍，p <0.01）明显更大。总之，统计方法，生物图像信息学和机器学习的组合是用于分析微藻培养种群动态的自动化调查的有用方法。我们开发了一种随机森林模型，使用指纹矢量功能对活和死微藻种群的分布进行分类，其总体AUC为94.5％，准确度为82％。我们发现，与活细胞相比，死细胞的平均直径（1.4倍，p <0.01）和面积（1.5倍，p <0.01）明显更大。总之，统计方法，生物图像信息学和机器学习的组合是用于分析微藻培养种群动态的自动化调查的有用方法。