Classification models for the counting of cellular objects

The automated image analysis cell classification paradigm for estimating the proportion (P) of cells on a cytology slide containing objects of interest is presented. Automated cell counters based on image analysis offer a mechanized alternative to the tedious and time-consuming task of manually performing these counts. Several classification models for increasing the automated estimation accuracy of P are presented. The receiver operating characteristic (ROC) curve, as used in classical signal detection theory, provides the conceptual structure and mathematical foundation for the models. It is shown that simple formulations, using this theory, yield dynamic strategies that result in higher cellular object classification accuracy than the classical one-threshold signal detection model. Moreover, these strategies can be implemented in a manner that satisfies the specific application constraints. An application involving the estimation of the proportion of various immunologically labeled lymphocyte subpopulations illustrates the methodology