Breast cancer is the second leading cause of cancer deaths in women today. Mammography is currently the primary method of early detection. But research has shown that many cases missed by mammography can be detected in breast MRI. Breast MRI is more difficult to interpret than mammography, because it generates much more data. Also, because it is a non-standard modality, there are fewer people qualified to use it for diagnosis. Our primary motivation is to increase the classification specificity for readers by using automatic classification system. We propose a method for finding significant regions in the data, and classifying them into malignant and benign. In this pilot study, our method first searches the breast tissue for contrast enhanced regions, then computes a difference image by subtracting the pre-contrast and post-contrast images. Two levels of thresholding are then applied. These thresholds are used to identify suspicious lesions and separate them from the background. After thresholding, 3D connected components analysis is used to label the enhanced lesions, and lesion shape based features are computed and used as inputs to the classifier. The last step is to use a backpropagation neural network (BNN) to classifying the labeled regions into benign and malignant, using the biopsy results as the gold standard. Our preliminary results show an area under the ROC curve for the testing stage equals to 0.913. This result illustrates the promise of using BNN as a physician's assistant for breast MRI classification.