The assignment of a cellular identity to individual pixels in microscopy images is a key technical challenge for many live-cell experiments. Traditional approaches to this image segmentation problem have relied on standard computer vision techniques, such as thresholding, morphological operations, and the watershed transform. While these approaches have enabled the analysis of numerous experiments, they are limited in their robustness and in applicability. Here, we show that deep convolutional neural networks, a supervised machine learning method, can robustly segment the cytoplasms of individual bacterial and mammalian cells. This approach automates the analysis of thousands of bacterial cells and leads to more accurate quantification of localization based fluorescent reporters in mammalian cells. In addition, this approach can also simultaneously segment and identify different mammalian cell types in co-cultures. Deep convolutional neural networks have had a transformative impact on the problem of image classification, and we anticipate that they will have a similar impact for live-cell imaging experiments.

Visit our webpage at http://covertlab.github.io/DeepCell