Performance of Triple-Modality CADx on Breast Cancer Diagnostic Classification

The purpose of this study is to evaluate the potential of computer-aided diagnosis (CADx) methods utilizing three breast imaging modalities: full-field digital mammography (FFDM), sonography, and dynamic contrast-enhanced magnetic resonance imaging (DCE-MRI) for breast lesion classification Three separate databases for each modality were retrospectively organized: FFDM (255 malignant lesions, 177 benign lesions), ultrasound (968 malignant lesions, 158 benign lesions), and DCE-MRI (347 malignant lesions, 129 benign lesions) From these single-modality databases, three dual-modality databases were constructed as well as a triple-modality database (31 malignant lesions, 17 benign lesions) Our computerized analysis methods consisted of several steps: (1) automatic lesion segmentation; (2) automatic feature extraction; (3) automatic feature selection; (4) merging of selected features into a probability of malignancy Stepwise linear discriminant analysis using a Wilks lambda cost function in a leave-one-lesion-out method was used for feature selection The selected features were merged using a Bayesian artificial neural network (BANN) with a leave-one-lesion-out method The classification performance was assessed using receiver-operating characteristics (ROC) analysis Results showed that the computerized analysis of breast lesions using image information from all three modalities yielded an AUC of 0.95±0.03 The observed trend of increasing performance as information from more modalities is included in the classifier indicates that the use of all three modalities can potentially improve the diagnostic classification of CADx.