Nucleus Segmentation Using Gaussian Mixture based Shape Models.

Abstract

We identify cells in microscopy images with stained nuclei, using the following process: Candidate seeds for nuclei are identified as extrema in a Laplacian-of-Gaussian space, and weak candidates are eliminated from clusters obtained by ellipse fitting; a region of interest for each nucleus is then defined by combining local and global thresholding; and these regions are repeatedly merged and split by modeling the shape of a nucleus and measuring the roughness of the shared boundaries connected nuclei. This method showed superior abilities to detect the nucleus regions and to split the boundaries of connected nuclei. Our experiments show higher scores in comparison with five other techniques in terms of eight evaluation metrics.