A Unified Segmentation Method for Detecting Subcellular Compartments in Immunofluorescently Labeled Tissue Images

We present a unified segmentation framework for detecting both membrane and nuclei structures in microscopy images of immunofluorescently labeled histological tissue sections. The non-parametric method presented can handle arbitrary mixtures of bloband ridge-like structures, which is commonly found in tissue images. The algorithm iteratively estimates the empirical likelihood functions of curvature and intensity based features of nuclei and membrane structures. The method was compared to manual segmentation, binary thresholding and watershed, and achieved up to 97.1% sensitivity and 94.4% specificity compared to manual segmentation. Scores measuring target protein expressions in each of the segmented subcellular compartments were also computed. For estrogen receptor (ER), the automatically obtained expression scores achieved 96% sensitivity and 90% specificity compared to manual assessment.