Classification of breast lesions in dynamic contrast-enhanced MR images

Abstract

In recent years, the development of computer-aided diagnosis (CAD) for breast MR image (MRI) has been a big challenge. Usually multiple layer perceptron (MLP) was used for classification of breast MRI lesions. Fuzzy technique can integrate human expert's knowledge into the system and integrating it with artificial neural network (ANN) could provide us with more intelligent systems. Therefore, in this work, a three-layer feed-forward MLP classifier and a four-layer feed-forward fuzzy neural network (FNN) classifier were used separately to compare their diagnostic performance in discrimination between malignant and benign breast lesions. This work included 40 (23 malignant and 17 benign) histopathologically proven lesions and the steps of this work were as follows: region of interest (ROI) selection, fuzzy c-means (FCM) segmentation, some morphological feature extraction, MLP and FNN classifications, Receiver Operating Characteristic (ROC) analysis. The results showed FNN classifier has a better diagnostic performance than MLP classifier in discrimination between malignant and benign lesions, because FNN classifier has a greater accuracy and area under the receiver operating characteristic curve (AUC) than MLP classifier, and also at the similar sensitivity, FNN classifier has a greater specificity than MLP classifier. This indicates FNN could provide us with good performance in discrimination between malignant and benign breast lesions which can lead to more powerful breast MRI CADs.