Whole cell segmentation of dense cell cultures in transmitted light images by deep learning.

Abstract number
European Microscopy Congress 2020
Corresponding Email
[email protected]
DHA.1 - Deep learning for analysis and interpretation of microscopy imaging data
dr Christoffel Dinant (1)
1. Danish Cancer Society Research Center

deep learning, segmentation, U-net

Abstract text

Here we present a deep learning-based image segmentation algorithm that takes transmitted light images and returns whole cell masks. 

Quantitative microscopy of cells universally requires accurate segmentation of the images into single objects. Image analysis can then be performed on a per-object basis, for example we can measure the fluorescence intensity inside or in an area around each cell nucleus. For this purpose a DAPI-stained cell nucleus can usually be segmented by conventional intensity threshold-based image processing tools because of strong homogeneity in shape and size parameters and staining intensities, but in other cases we might want to segment other objects, e.g. whole cells, for which these parameters can vary drastically within and between samples dependent on cell type, confluence and treatment. In addition we would often prefer to use a transmitted light image for the segmentation step in order to free up as many fluorescence channels as we have available for simultaneous analysis.

To create ground truth image masks we made six U2OS cell cultures containing two each of four cytoplasmic fluorescence markers with emission in the blue, green, red and far-red spectra. These six cell cultures were thoroughly mixed together and plated unto coverslips at increasing cell densities after which we acquired images of the four above-mentioned fluorescence channels plus transmitted light. Because the vast majority of cells have neighbors with a different cytoplasmic color combination we can now create whole-cell binary segmentation masks just by thresholding on each of the six fluorescence channel combinations. These binary masks were used as the ground truth for training a U-net deep neural network with the transmitted light images as input. Results will be presented at the conference.

 Accurate whole-cell segmentation of transmitted light images under many different culturing conditions will dramatically improve the precision of quantitative analysis of fluorescence signals from cytoplasmic organelles and other sub-cellular structures.