Combining multiple mammographic views for computer aided diagnosis

Abstract

The agreement in mammographic lesion features from two projections of the breast was studied for human extracted BI-RADS" and computer extracted image texture features and cases with low agreement were identified. Factors affecting agreement between the two views were isolated and agreement for the two types of features was compared. Agreement in BI-RADS" features was higher than that for texture features, however agreement for both types of features was affected by the institutional source of data. Statistical classifiers for distinguishing benign from malignant lesions were designed using BI-RADS" features from multiple views and the effect of including patient age was investigated. The highest AUC ± SD value of 0.927 ± 0.026 was obtained for a classifier that used features from both the MLO and CC views as well as patient age as features.