Micromass image segmentation

Brightfield images of face and limb buds cells (Fig. ​(Fig.1a)1a) at 20 days in normal micromass culture were analysed with Matlab® (2017b) to extract condensation shape descriptors. Images were first converted to grayscale (Fig. ​(Fig.1b),1b), preprocessed to enhance contrast by contrast-limited adaptive histogram equalization (CLAHE) and low-pass filtered to remove constant power additive noise before being binarized (Fig. ​(Fig.1c)1c) with Otsu’s method. Morphological opening was performed to remove any small white noises in the image, and morphological closing to remove any small holes in the object. All connected components that had fewer than 24 pixels were removed and structures that were lighter than their surroundings and connected to the image border were suppressed. The images were segmented by a watershed transformation (Fig. ​(Fig.1d)1d) and a distance transform was used as segmentation function to split out the regions. Watershed transform is known for its tendency to “oversegment” an image because each local minimum, no matter how small, becomes a “catchment basin”. For this reason we (i) filtered out tiny local minima using imextendedmin (ii) modified the distance transform so that no minima occured at the filtered-out locations. This procedure is called “minima imposition” and was implemented via the function imimposemin.

With watershed segmentation, single condensations were identified. The properties of the image regions were directly or indirectly calculated through the regionprops function (Table ​(Table11).

Linear Discriminant Analysis (LDA) between face and limb buds condensation shape descriptors was performed (Fig. ​(Fig.1e).1e). Shape descriptors were extracted from experiments with three technical repeats with three biological replicate for each one. The obtained values were standardized for all of the predictors (shape descriptors) and the regression model was fitted with the fitcdiscr function of Matlab®.