HUMAN INDUCED PLURIPOTENT STEM CELL REGION DETECTION IN BRIGHT-FIELD MICROSCOPY IMAGES USING CONVOLUTIONAL NEURAL NETWORKS

Abstract

Human induced pluripotent stem (iPS) cells represent an ideal source for patient specific cell-based regenerative medicine. For practical uses of iPS cells, large-scale, cost- and time-effective production of fully reprogrammed iPS cells from a number of patients should be achieved. To achieve this goal, culture protocols for inducing iPS cells as well as methods for selecting fully reprogrammed iPS cells in a mixture of cells which are still in reprogramming and non-iPS differentiated cells, should be improved. This paper proposes a convolutional neural network (CNN) structure to classify a bright-field microscopy image as respective probability images. Each probability image represents regions of differentiated cells, fully reprogrammed iPS cells or cells still in reprogramming, respectively. The CNN classifier was trained by multiple types of image patches which represent differentiated, reprogramming and reprogrammed iPS cells, etc. Classification of an image containing the confirmed iPS cells by the trained CNN classifier shows that high classification accuracy can be achieved. Classifications of sets of time-lapse microscopy images show that growth and transition from CD34[Formula: see text] human cord blood cells through reprogramming to reprogrammed iPS cells can be visualized and quantitatively analyzed by the output time-series probability images. These experiment results show our CNN structure yields a potential tool to detect the differentiated cells that possibly undergo reprogramming to iPS cells for screening reagents or culture conditions in human iPS induction, and ultimately further understand the ideal culturing conditions for practical use in regenerative medicine.