Automatic Single-Cell Discrimination by Cellular Appearance using Convolutional Neural Network

Morphological images of cells contain extensive information, which help biologists to infer the type and state of cells to some degree based on their morphology. Convolutional Neural Network, a neural network architecture, is a powerful tool used for image recognition. However, whether it can be used to classify cells on the basis of their morphology remains unclear. In this study, we demonstrate that 10 different hematopoietic tumor cell lines with similar morphologies that biologists find difficult to distinguish can be classified with >90% accuracy using only their bright-field images when analyzed using a Convolutional Neural Network. This novel and simple system using bright-field images of cells could be a powerful analytical tool for cell type discrimination and could also be applied to the clinical diagnoses of hematopoietic tumors.