Comparison of Segmentation Methods for Cytometric Assay

In this paper we compare two segmentation techniques, a parameter-free variant of the watershed algorithm and a new robust method based on adaptive contours. To test them, we use a set of fluorescence in-situ hybridisation images of leucocyte nuclei: these images were obtained in realistic conditions of mass screening, and do not meet the prerequisites of usual segmentation methods. We present the two methods and describe them step by step. We then compare the results of the segmentation of the test images, showing over a hundred nuclei, and detail the incorrect segmentations. The contour-based method outperforms the watershed as regards correct segmentations and over-segmentations, but performs more under-segmentations. We analyse the origins of the incorrect segmentations from the two methods, and discuss possible improvements.