Image Segmentation of Cell Nuclei based on Classification in the Color Space

Abstract

Many developments in the field of image processing and analysis have been motivated and driven by applications from microscopy. Unfortunately, today hardly any general applicable set of image processing methods exist to support biomedical experts with a fully or partially automated analysis of micrographs to support and strengthen his experiments. Hence, currently much image analysis and interpretation in this field has to be done manually. Otherwise, a background in image processing is needed to adopt available image analysis tools to the desired task, or to write scripts in image-processing toolboxes.

Thus, using cervical nuclei as a first example, a novel image processing scheme is suggested to bridge the so-called “semantical gap” between the analytical question of the biomedical experts on one side and the required set of image processing procedures on the other. Within this approach, the biomedical expert interactively annotates the background and foreground pixels (nuclei) on a small subset of micrographs of cervical cells. Using this information, the image processing system can now automatically find the most optimal classifier to separate these two sets of pixels in color space. Within this approach machine-learning algorithms, namely K-Nearest Neighbor, KStar, Gaussian Mixture Models, KMeans and Expectation-Maximation have been used and evaluated for the task to separate fore- and background pixels in the color space to yield an automated segmentation approach for cervical nuclei. Thus, the suggested approach is able to “learn” the required segmentation task from the biomedical experts by interactively training the system on a small subset of images. After the self-organization and optimization procedure, the system is capable to apply the learned analysis mechanisms to other micrographs of cervical cells.