Radial symmetries based decomposition of cell clusters in binary and gray level images

The segmentation of structures of complex cytological and histological images is a necessary intermediate step for image analysis that give rise to binary images. In many cases these binary images can be rather far away from a subsequent object specific quantification because biological structures digitized by optoelectronic devices may situated close together so that they appear as one fused object in the projective image. Such fusions of objects may become complex so that large clusters of biological structures emerge. To quantify individual objects of a cluster they must be separated. The shape, size and intensity variation of cells in complex organs like the brain may breed planar configurations that can be splitted only inadequate by common techniques, e.g., watershed separation or basic morphological processing of images. Considering iteratively object contours suitable features of saliency can be accumulated that give rise to markers of singular objects. Such significant markers may drive a separation process more effective than common approaches. The determination of markers by an iterative method should be scale, translation and rotation invariant and robust with regard to noise due to the variability of biological specimen. We realize a technique that splits cell clumps consisting of different cell sizes and shapes into meaningful parts. The multiscale method applied here is based on the analysis of the contour shape and the object area by iterative voting using oriented kernels. These cone-shaped kernels vote iteratively for the local center of mass of the components of an aggregation. The voting is performed along the gradient of the distance transformation of the binarized image of aggregates. Iterative voting is initialized by voting along the gradient direction where at each iteration the voting direction and shape of the kernel is refined, resp. the kernel topography is refined and reoriented iteratively. It turned out that the kernel topography is unique because it votes for the most likely set of grid points where the gravity center of an individual cluster component may be located. Furthermore, a new procedure is realized to use the local intensities of aggregations for kernel voting. The last voted iteration provides gravitation centers, resp. centers of mass of the clumped cells. These are extracted and used as markers to determine individual cell boundaries by a marker based watershed postprocessing. The subject of this paper is to highlight the basic algorithm of iterative kernel voting and expanding it to process intensities within clusters as well as contour information. The approach is applied to synthetic images that were modified systematically with regard to object topology. Natural aggregates of cells at the light microscopic level and cell clusters derived from high resolution flat bed scanning were splitted. In addition to these examples images from a benchmark databases were investigated. The splittings generated by the iterative voting approach were compared with expected splittings of test persons and with results of the watershed method. Especially the gray level based iterative voting method provides superior results for cell cluster separation in comparison to the watershed procedure.