Improvement of Performance to Discriminate Malignant Tumors from Normal Tissue on Mammograms by Feature Selection and Evaluation of Feature Selection Criteria

摘要

This paper reports the results of improving the performance of a previous system for discriminating malignant tumors by selecting a small number of features effective for discriminating malignant tumors from normal tissue on mammograms from 39 features. Specifically, the efficacy of the new system is evaluated by comparing the results with the case of using the previous 9 features. The paper also reports on the results of comparing the areas A{sub}z under ROC curves using Mahalanobis distance ratios, the areas A{sub}z under ROC curves based on the maximum likelihood method, and the average recognition rates based on the nearest-neighbor decision rule.