Characterization of chromatin texture by contour complexity for cancer cell classification

The purpose of this study is to investigate a new technique for image-based cancer cell classification and provide a more quantitative and objective characterization method for a diagnosis, which currently relies on qualitative and empirical judgment of pathologists. For this, a new method for chromatin texture characterization employing a new feature, contour complexity, is proposed and evaluated using nuclear images obtained from paraffin-wax embedded sections of human breast cancer on slides. The proposed feature is calculated on the basis of a contour length of nucleus obtained by setting different threshold values of intensity for a grayscale image, and it is a quantitative measure of chromatin texture. An expectation-maximization (EM) algorithm-based segmentation and an effective initial parameter search method for EM are used for the automatic calculation of the feature. The results for breast cancer cell detection showed that the average contour complexity value for malignant cells (19.6plusmn4.1) is found to be significantly greater (p < 10^-6, Kolmogorov-Smirnov test) than that of benign cells (0.35plusmn0.17). By the comparison with the conventional fractal dimension approach, it is shown that the proposed feature is much more sensitive feature than the fractal dimension for the individual cancer cell detection.