Automated Counting of Cancer Cells by Ensembling Deep Features
Automated Counting of Cancer Cells by Ensembling Deep Features
复制标题
DOI：
10.3390/cells8091019
复制
发表时间：
2019-09-01
期刊：
影响因子：
6
通讯作者：
Hu, Pingzhao
中科院分区：
文献类型：
--
作者：
Liu, Qian;Junker, Anna;Hu, Pingzhao
高容量和高通量数字显微镜在生物实验和临床实践中产生了大量的图像集。自动图像分析技术（例如细胞计数）的需求很高。在这里，细胞计数被视为使用深度学习模型提取的图像特征（表型）的回归问题。开发了三种深度卷积神经网络模型，以端到端的方式将图像特征回归到其细胞计数。从理论上讲，集合成像表型应比单一类型的成像表型具有更好的代表性。我们实现了这一想法，通过整合两种类型的成像表型（点密度图和前景掩模）提取的两个自动编码器和回归的集合成像表型细胞计数之后。使用两个公开的合成显微图像数据集来训练和测试所提出的模型。采用均方根误差、平均绝对误差、平均绝对百分比误差和皮尔逊相关系数来评价模型的性能。训练好的模型也被应用于预测生物实验中获得的真实的显微图像的癌细胞计数，以评估两个结直肠癌相关基因的作用。与基于单一成像特征的模型相比，通过集成深度成像特征的模型在更小的误差和更大的相关性方面表现出更好的性能。总的来说，所有模型的预测都显示出与真实细胞计数的高度相关性。基于集成的模型集成了高水平的成像表型，以改善对来自高内容和高通量显微图像的细胞计数的估计。
High-content and high-throughput digital microscopes have generated large image sets in biological experiments and clinical practice. Automatic image analysis techniques, such as cell counting, are in high demand. Here, cell counting was treated as a regression problem using image features (phenotypes) extracted by deep learning models. Three deep convolutional neural network models were developed to regress image features to their cell counts in an end-to-end way. Theoretically, ensembling imaging phenotypes should have better representative ability than a single type of imaging phenotype. We implemented this idea by integrating two types of imaging phenotypes (dot density map and foreground mask) extracted by two autoencoders and regressing the ensembled imaging phenotypes to cell counts afterwards. Two publicly available datasets with synthetic microscopic images were used to train and test the proposed models. Root mean square error, mean absolute error, mean absolute percent error, and Pearson correlation were applied to evaluate the models' performance. The well-trained models were also applied to predict the cancer cell counts of real microscopic images acquired in a biological experiment to evaluate the roles of two colorectal-cancer-related genes. The proposed model by ensembling deep imaging features showed better performance in terms of smaller errors and larger correlations than those based on a single type of imaging feature. Overall, all models' predictions showed a high correlation with the true cell counts. The ensembling-based model integrated high-level imaging phenotypes to improve the estimation of cell counts from high-content and high-throughput microscopic images.