Optimizing cervical cell classifiers.

Abstract

In an automated prescreening system where a cell classifier and a specimen classifier operate in cascade, the false-positive and false-negative error rates of each classifier can be traded off to obtain the best overall performance. It is usually desirable to keep the specimen false-negative rate below the false-positive rate. An analysis of the classifier cascade shows that, in contrast, the cell classifier should have its false-positive rate much lower than its false-negative rate. A procedure is presented for selecting the best operating point on the ROC curve of the cell classifier. This minimizes the sample size required to achieve prescribed specimen error rates.