Reliability Assessment of Ensemble Classifiers: Application in Mammography

摘要

In classifier ensembles predictions of different classifiers regarding a query are combined into one final decision. It was previously shown that using ensemble techniques can significantly improve classification performance. In this study we build upon this result and propose to use variability in the predictions of classifiers contributing to the final decision as an indicator of its reliability. The study hypothesis is tested with respect to previously proposed information-theoretic computer-aided decision (IT-CAD) system for detection of masses in mammograms. A database of 1820 regions of interest (ROIs) extracted from digital database of screening mammography (DDSM) is used. Experimental results show that the proposed reliability assessment successfully identifies decisions that can not be trusted. Further, a low correlation between reliability and the classifier output is noted. This opens a possibility of combining reliability and ensemble output into one improved decision.