In order to improve the separation between abnormal cells and noncellular artifacts in the CERVIFIP automated cervical cytology prescreening system, 22 different object texture features were investigated. The features were all statistical parameters of the pixel density histograms or one-dimensional filtered values of central and border regions of the object images. The features were calculated for 231 images (100 cells and 131 artifacts) detected as Suspect Cells by the current CERVIFIP and were then tested in hierarchical and linear discriminant classifiers. After selecting the two best features for use in a hierarchical classifier, 83% correct classification was achieved. One of these features was specifically designed to remove poorly focused objects. With maximum likelihood discrimination using all 22 features, an overall correct classification rate of 90% was obtained.
Key terms: Image analysis, pattern recognition features, automated prescreening, cewical cytology chromatin, texture, artifactsThe problem of rejecting noncellular artifacts is one of very great practical importance in automated cytometry. For example, a major difficulty in cervical cytology automation is that of rejecting false-alarm signals ("false positives") in normal specimens caused by artifacts such as overlapping cell pairs, leukocyte clusters, dust particles, etc. Already, several techniques have been developed for this purpose based upon the analysis of global object features (e.g. area, density) (14,19), object shape (6,14,16-18,24,25), object texture (61, and object colour (6). Nevertheless, it has been found in system trials that a significant number of artifacts are not detected using classification based upon these features, so that falsepositive objects are still a major factor in specimen misclassification. Improvements in features for artifact rejection should therefore result in an increased eficiency of such systems.The analysis of cellular texture is one of the major areas of research in the extraction of information from cell images. Several different approaches have emerged, including the use of co-occurence (transition) matices (23), grain analysis (13), one-and two-dimensional filtering (lo), run-length texture analysis (8), and statistical parameters (1). In quantitative cytology, the analysis of nuclear texture has been particularly successful in separating normal and malignant specimens, even in apparently normal cells, (3,4), because texture measurements on an appropriately prepared and stained cell nucleus reflect the local chromatin arrangement, which is one of the most powerful indicators of cell function and hence of cell pathology.One type of instrumentation that has found application in the field of automated cytometry is that of the continuous motion imaging (CMD system, in which the specimen is moved continuously under a one-dimensional scanner (e.g., linear CCD diode array) (2,12,15).These systems possess several desirable properties for cervical cytology automation; they are fast, and they have potentia...
