A least upper bound for the average classification accuracy of multiple observers

摘要

The purpose of this paper is to present an upper bound equation for the average classification accuracy of multiple observers subjectively classifying a finite collection of objects. Research in pulmonary cytopathology has provided advances in the early detection of squamous cell carcinoma of the lung. Current efforts in this field include an attempt to accurately detect and identify pre-cancerous states of the lung. The upper bound equation in this paper resulted from an analysis to define training data and evaluation techniques for automated computer classification of pre-cancerous states of lung cells. The result can be applied to any multiple observer classification problem. In the absence of “truth” information, the result provides a least upper bound for the average classification accuracy of the observers.