Automated Classification of Analysis- and Reference Cells for Cancer Diagnostics in Microscopic Images of Epithelial Cells from the Oral Mucosa

Abstract

To get the best possible chance of healing, cancer has to be detected as early as possible. As cancer starts within a single cell, cytopathological methods offer the possibility of early detection. One such method is standardized DNA image cytometry. For this, the diagnostically relevant cells have to be found within each specimen, which is currently performed manually. Since this is a time-consuming process, a preselection of diagnostically relevant cells has to be performed automatically. For specimens of the oral mucosa this involves distinguishing between undoubted healthy epithelial cells and possibly cancerous epithelial cells. Based on cell images from a brightfield light microscope, a set of morphological and textural features was implemented. To identify highly distinctive feature subsets the sequential forward floating search method is used. For these feature sets k-nearest neighbor and fuzzy k-nearest neighbor classifiers as well as support vector machines were trained. On a validation set of  cells it could be shown that normal and possibly cancerous cells can be distinguished at overall rates above 95.5 % for different classifiers, enabling us to choose the support vector machine with a set of two features only as the classifier with the lowest computational costs.