Simultaneous Segmentation of Cell and Nucleus in Schizosaccharomyces pombe Images with Focus Gradient

Schizosaccharomyces pombe shares many genes and proteins with humans and is a good model for chromosome behavior and DNA dynamics, which can be analyzed by visualizing the behavior of fluorescently tagged proteins in vivo [1]. However, performing a genome-wide screen for changes in such proteins requires developing methods that automate analysis of multiple images. The first step requires robust segmentation of the cell and the most distinguishable compartments (the nucleus) from images with varying focus conditions and qualities. We developed a segmentation system that can segment transmitted illumination images with focus gradient and varying contrast, and extract cell and nucleus boundaries. Global and locally adaptive corrections for focus gradient are applied to the image to accurately detect cell membrane and cytoplasm pixels. We use the gradient vector flow snake model [2] to segment individual cells, using a novel edge map based on detected cell membrane. We applied our system to multi-channel images of S. pombe, the whole data set contains about 4000 mutant genotypes each with at least three sets of transmitted illumination (bright field), Rad52-YFP and RPA-CFP images. Our system is able to correctly segment a majority of nuclei and cells in almost all images of sufficient quality, and performance is consistent over a wide variety of focus distance, field brightness, relative contrast and phenotypic characteristics. A quantitative evaluation is also performed using a set of hand produced gold standard segmentations of pombe cells, representing different image acquisition conditions and quality. We evaluated the percentage of cells detected, the accuracy of the final snake contours. The whole set of 60 gold standard images contain a total of 14,926 pombe cells, averaging about 249 cells per image, of which 97.5% were detected by nucleus segmentation and pixel classification of cell interior, and 89.0% were accurately segmented (defined as less than 10% pixel mismatch). Our system generated a total of 16,631 snake contours, of which 88.3% are true positives, the rest being false detections, incorrect merging or partial segmentation. After erroneous cell contours are removed by an automatic contour validation classifier, the remaining cell contours contain 98.3% true positives, this shows that although our system has a modest segmentation accuracy, the final cell contours generated is very reliable overall. For large scale high-throughput applications with huge amounts of data, in order to minimize the need for human intervention, the high reliability and robustness achieved by our system is very valuable. We have also compared with recent methods [3], and our method. In conclusion we have developed a multi-channel cell and nucleus segmentation system for S. pombe cells that uses nucleus protein fluorescence to correct for varying focus and contrast in the transmitted illumination image, combined with active contour segmentation and robust automatic contour validation. This system can be applied to similar light microscopy images where some fluorescence signal within the cell nucleus or cytoplasm is provided, and can in principle be extended to deal with multiple cell types and image modalities.