Modeling Structural Dissimilarity Based On Shape Embodiment For Cell Segmentation

Accurate cell segmentation is one of the critical, yet challenging problems in microscopy images due to ambiguous boundaries as well as a wide variation of shapes and sizes of cells. Even though a munber of existing methods have achieved decent results for cell segmentation, boundary vagueness between adjoining cells tended to cause generation of perceptually inaccurate segmentation of stained nuclei. We propose a segmentation method of cells based on structural dissimilarity between embodied and imaged cells. From assumption that the shape of the region of adjoining cells follows a 2D Gaussian mixture model, the cell region is divided by an expectation-maximization method. The lowest structural dissimilarity using embodied cells decides on the number of components of the 2D Gaussian mixture model. The region of interest is extracted by implementation of both global and local thresholdings, which performs binarization of the local image with a seed at the center, where the seed is obtained by the maximally stable extremal regions. Our approach presented considerably higher evaluation scores compared with other five existing methods in terms of both accuracies of region of interest (ROI) detection and boundary discrimination.