The performance of a cell recognition system on unknown data is often estimated in terms of its error rates on a test set. This paper reports on the investigation of methods for producing estimates of error rates in cervical cell classification. One method, often used, which estimates error rates is the holdout method in which a portion of the available data is used to train the classifier and the remainder is used as a test set. This method of error rate estimation is much too unreliable for use in performance evaluation or comparison of cervical cells classifiers. The advantages of the holdout method are obtained in a estimate with lower variance by averaging the estimates from many randomly partitioned holdout experiments. Classification performance curves calculated using other methods are given for several classification schemes used to classify 1500 cervical cells.
ASJC Scopus subject areas