Discriminant analysis for classification of murine melanomas and human cervical epithelial cells.

Computer analysis of cell images offers many advantages over routine visual examination. It leads to quantitative and accurate detection of subvisual information and provides reproducible measures so that objective decisions in cancer diagnosis become possible. Such diagnostic decisions usually follow partly from a classification process. In this paper two multivariate discriminant analysis methods--namely, linear discriminant analysis (LDA) and quadratic discriminant analysis (QDA)--are presented. LDA and QDA were used to classify cytologic data based on some morphodensitometric measurements. The cytologic data constituted two samples, one representing B16 cell lines and the other including three types of normal human cervical epithelial cells. LDA and QDA were assessed both individually and in comparison to each other, mainly on the basis of the rate of correct classification and robustness. The measurements extracted from the cytologic data employed were shown to be stable and consistent. The statistical results obtained from experiments on cervical cells look particularly promising and encouraging for future work. It has also been shown in this study that the classification techniques employed are valid and that LDA performed almost as well as QDA.