Polar Space Based Shape Averaging for Star-shaped Biological Objects

Abstract

In this paper, we propose an averaging method for expert segmentation proposals of microbial organisms, resulting in a smooth, naturally looking segmentation ground truth. The approach exploits a geometrical property of the majority of the organisms - star-shapedness - and is based on contour averaging in polar space. It is robust and computationally efficient, where robustness is due to the absence of tuneable parameters. Moreover, the algorithm preserves the uncertainty (in terms of the standard deviation) of the experts' opinion, which allows to introduce an uncertainty-aware metric for estimation of the segmentation quality. This metric emphasizes the influence of ground truth regions with low variance. We study the performance of the proposed averaging method on time-lapse microscopy data of Corynebacterium glutamicum and the uncertainty-aware metric on synthetic data.