On the Decomposition of Cell Clusters

Successful segmentation of a multilevel to a bilevel microscopic cell image rather frequently gives rise to touching objects which need to be separated in order to perform object specific measurements. The standard approach of dealing with this problem is a watershed decomposition of gradient, distance or low pass filtered transforms. However, if cell clustering is excessive, the cell size varies and cells have various shapes that are different from circles the watershed approaches produce unsatisfying results.We found a technique that splits cell clumps into meaningful parts. Since this method is based on the analysis of contour curvature on the scale space of Fourier coefficients relevant dominant points can be recognized. Based on an optimized heuristic approach pairs of these dominant points are recursively matched since splitted objects do not possess concavities respectively intrusions anymore. The advantages of this approach are (i) the independence of cell shapes which are clumped, (ii) the consideration of holes or background intensities within objects, (iii) the robustness in terms of convergence and a few parameters only to adapt to other families of decomposition problems.The objective of this contribution is to explain the algorithm, show its results using different examples from benchmark databases, self generated images and complex configurations of cell images.