Abstract
Clumps of adenocarcinoma are found in about 10% of cytologically positive findings in mass screening of uterine cancer. Since they are not easily dispersed into individual cells, prescreening machines designed to examine individual cells fail to detect their presence. This paper presents two feature extraction methods, global and local, for the recognition of such cell clumps. The global feature extraction method is designed for prescreening machines, to extract two features utilizing digital image processing. One feature is an Euler number. A thresholded binary image is shrunk and propagated, so that only large objects corresponding to cell clumps remain. The Euler number of the remaining objects is a good measure of overlap of nuclei, one of morphological features of adenocarcinoma cell clumps. Another extracted feature is a correlation coefficient of binary images obtained at two different threshold levels. This value indicates the interior structure of the object, such as the amount of cytoplasm. Experiments using 79 samples, consisting of adenocarcinoma cell clumps, normal cell clumps, white blood cell clumps, and dust particles, yielded good results : i.e., false negative proved zero and false positive rate was 11%. The local feature extraction method is intended to examine the correspondence between morphological features described by medical doctors and the parameters extracted by machine. In this method each individual nucleus of a cell clump is examined. The center points and the radii of nuclei are computed by the process based mostly on the distance transformation for binary images. The variation in the distribution of radius and minimum distance among nuclei shows a good correspondence with the morphological features, i. e., irregularity of nuclear size, and irregularity of distance among nuclei.
