Breast cancer is now considered as one of the leading causes of death among women all over the world. It is broadly accepted that ultrasound imaging is an important and frequently used tool for breast cancer diagnosis. In this paper, we propose a novel computer-aided diagnosis scheme for breast lesions classification. In this scheme, the sonographic breast images are first used to produce Breast Imaging Reporting and Data System (BI-RADS) lexicon based feature scoring data. Biclustering mining is then used as a powerful tool to discover the effective local diagnosis patterns in training data, and those found biclusters are utilized to generate hidden features as new input data. Finally the back-propagation (BP) neural network algorithm is applied to produce an efficient classifier for recognizing benign and malignant breast tumors. Our experimental results show that the proposed method yielded good prediction performance, indicating its interesting potential in clinical applications.