On the possibility of reliably constructing a decision support system for the cytodiagnosis of breast cancer

Abstract

We evaluate the performance of three Bayesian network classifiers as decision support system in the cytodiagnosis of breast cancer. In order to test their performance thoroughly, we use two real-world databases containing 692 cases collected by a single observer and 322 cases collected by multiple observers respectively. Surprisingly enough, these classifiers generalize well only in the former dataset. In the case of the latter one, the results given by such procedures have a considerable reduction in the sensitivity and PV- tests. These results suggest that different observers see different things: a problem known as interobserver variability. Thus, it is necessary to carry out more tests for identifying the cause of this subjectivity. © 2007 Springer-Verlag Berlin Heidelberg.