Stabilized binary hierarchic classifier in cytopathologic diagnosis

A binary tree classifier (BTC) algorithm for computer-assisted cell image analysis has been developed that overcomes the problem of overtraining due to inadequate sample size/dimensionality ratio at the higher-order nodes of a hierarchic decision structure. Provisions have been introduced that ensure that decision rules created at each node are based on samples representative of the subpopulation routed to the node. These provisions eliminate problems caused by truncation effects resulting from the application of decision rules at preceding decision nodes. The BTC performs better than do single-stage classifiers in situations where the categories' mean vectors are not well separated and no equality of covariance matrices exists. In applications in which noticeable deterioration of classifier performance on test-set data is common, the classification success rate of the BTC algorithm is not statistically significantly different between the training-set and test-set data.