Abstract
Multicellular organisms develop from the fertilized eggs through numbers of cell divisions. While the pattern of cell division has specific rules for healthy development, some fluctuations for this pattern are allowable. In order to uncover this robust mechanism of development, the position of cells in each embryo must be analyzed quantitatively. In the embryonic developmental biology, various studies are attempted to acquire the quantitative criteria from time-series three-dimensional microscopic images by image analysis such as segmentation, to understand the mechanism of development. Unfortunately, due to the fact of inaccuracy of nuclei detection and segmentation, it is a hard task to evaluate an embryo quantitatively from bioimages automatically. Based on these demands of quantitative analysis, we developed QCA Net, which accurately performs nuclear segmentation of three-dimensional fluorescence microscopic images for early-stage mouse embryos. QCA Net is based on Convolutional Neural Network. We trained QCA Net using a part of one early-stage mouse embryo. As a test, QCA Net performed segmentation of different 11 mouse embryos images. We succeeded in accurately acquiring the shape of the nucleus without fusion of nuclear regions. Besides, we achieved accurate extraction of the time-series data of nuclear number, volume, surface area, and center of gravity coordinates as the quantitative criteria of mouse development, from the segmentation images acquired by QCA Net. To our surprise, these results suggested that QCA Net recognized and distinguished the nucleus and a polar body formed in meiosis process. We consider that QCA Net can drastically contribute to performing segmentation of various bioimages in the embryonic developmental biology. The various quantitative criteria obtained from segmented images can uncover various unknown mechanisms of embryonic development, such as the robust mechanism of development.