Towards rapid cervical cancer diagnosis: automated detection and classification of pathologic cells in phase-contrast images.

In this paper, a combination of two methods based on texture analysis, contour grouping, and pattern recognition techniques is presented to detect and classify pathologic cells in cervical vaginal smears using the phase-contrast microscopy. The first method applies statistical geometrical features to detect image regions that contain epithelial cells and hide those regions with medium and contamination. Sequential forward floating selection was used to identify the most representative features. A shape of cells was identified by applying an active contour model supported by some postprocessing techniques. The second method applies edge detection, ridge following, contour grouping, and Fisher linear discriminant to detect abnormal nuclei. Evaluation of the algorithms' performance and comparison with alternative approaches show that both methods are reliable and, when combined, improve the classification. By presenting only images or their parts that are diagnostically important, the method unburdens a physician from massive and messy data. It also indicates abnormalities marking atypical nuclei and, in that sense, supports diagnosis of cervical cancer.