EPySeg: a coding-free solution for automated segmentation of epithelia using deep learning

Abstract

Epithelia are dynamic tissues that self-remodel during their development. During morphogenesis, the tissue-scale organization of epithelia is obtained through a sum of individual contributions of the cells constituting the tissue. Therefore, understanding any morphogenetic event first requires a thorough segmentation of its constituent cells. This task, however, usually involves extensive manual correction, even with semi-automated tools. Here, we present EPySeg, an open-source, coding-free software that uses deep learning to segment membrane-stained epithelial tissues automatically and very efficiently. EPySeg, which comes with a straightforward graphical user interface, can be used as a Python package on a local computer, or on the cloud via Google Colab for users not equipped with deep-learning compatible hardware. By substantially reducing human input in image segmentation, EPySeg accelerates and improves the characterization of epithelial tissues for all developmental biologists.

Keywords: Computer vision; Deep learning; Epithelia; Quantitative biology; Segmentation; Software.

Fig. 1.

EPySeg segmentation pipeline. An unseen image of cells labelled with a membrane marker is provided to the EPySeg pre-trained neural network. EPySeg produces seven outputs from it: five of them are watershed-like outputs, while the remaining two are watershed seeds. Those seven outputs are used to generate seven watershed masks. Upon thresholding the average of these seven masks, we obtain a refined mask.

Fig. 2.

EPySeg segmentation of unseen epithelium images. (A-D) EPySeg segmentation (red) overlaid on unseen images. (A) Segmentation of the Drosophila head epithelium, including ocelli, labelled with E-cadherin:GFP (greyscale). (B) Segmentation of the fourth leaf of a plasma membrane-labelled Arabidopsis thaliana plant at 7 days post germination (greyscale, UBQ10::acyl:tdTomato). (C) Segmentation of Phalloidin-labelled (greyscale) vertebrate dorsal pericardial wall epithelium (Cortes et al., 2018; Francou et al., 2017). (D) Segmentation of the Drosophila abdominal region surrounding a histoblast nest, labelled with E-cadherin:GFP (greyscale). Scale bars: 25 µm.