Utilization of Spatial Information for Segmentation of Cell Nuclei in Fluorescence Microscopy Image

Abstract

Numerous aspects of analyzing and quantifying fluorescence microscopy images rely on quantitative cell nucleus image analysis. Specifically, the basis for all automated cell image analysis in high-throughput applications is image segmentation. Semi-automatic and manual segmentation methods are tedious, require intensive labor, and suffer from inter-and intra-person variability. Therefore, automatic methods with the ability to deal with different cell types and image artifacts are required. The goal of this thesis is motivated by the fact that in cell nuclei image segmentation, the spatial relationship of the pixels can be utilized as an important characteristic that improves the performance of segmentation methods. Therefore, this thesis aims to use spatial information in cell nuclei image segmentation and proposes several implicit models under different assumptions. First, we assume that the local image data are Gaussian and propose an implicit model based on the Bayesian classification risk and anisotropic weighting scheme for fluorescence microscopy image segmentation. The proposed algorithm obtains the lowest MAD measure for all four experiments. The MAD values are approximately 18%, 7%, and 12% better than the state-of-the art methods we compared with for U20S cells, NIH3T3 cells, and Synthetic cells, respectively.