Towards quantifying the impact of cell boundary estimation on morphometric analysis for phenotypic screening

Computational analysis of appearance and morphology patterns of cells offers good potential to quantify the effect of candidate drug compounds on a cell population relative to the control in high throughput phenotypic screening. Extracting accurate morphometric information for each individual cell requires reliable detection and segmentation of object boundaries. For some cell lines and under certain culture conditions, images present cells with significant crowding, touching and boundary overlap. While very capable methods for segmenting touching and slightly overlapping cells have been developed, these can still only deal with a certain level of confluency. Here we propose to capture the uncertainty of the segmentation in the form of a confidence score and show how it can be incorporated into the downstream statistics. We show that by computing a mixture distribution of a morphometry metric with each component weighted by the associated segmentation confidence, we obtain a better approximation of the true underlining distribution of the given metric for a cell population compared with treating all segmented regions equally.