V. CONCLUSION AND FUTURE WORK
In this research report, we introduced the neural network classification model and ROC evaluation method to diagnose and classify malignant breast cancer and benign disease for breast biopsy outcome predictions based on the 830 clinical instances from the mammographic mass dataset. Our classification model, based on the two-stage back-propagation neural network classification approach, included both linear and nonlinear components for calculations as well as an input vector processor, adjustable learning rate controller, and training control center that allowed implementation of the iterative training processes. During each of the iterative training processes, the training model gradually increased the input data size to reuse the final trained weights from the previous iterative training stage as the initial weights for the next iterative training stage. Accordingly, the learning rate controller adjusted the learning rate. The proposed iterative training processes ensured that our model had a low SSE or MMSE. In order to obtain a highly accurate neural network diagnostic model, the proposed iterative training processes were especially useful for training our model when a large input of data and a large number of hidden neurons were present. Our research results showed that the neural network classification model had a specificity of 89.93% in diagnosing benign disease, a sensitivity of 89.33% in diagnosing malignant breast cancer, and an overall accuracy of 89.64% in diagnosing both malignant breast cancer and benign disease. An estimated area of the ROC curve for breast biopsy outcome predictions was 0.9626±0.0069. Therefore, our model along with mammography can provide highly accurate and consistent diagnoses for breast biopsy outcome predictions, allowing patients to bypass unnecessary surgical biopsies.