Classification of mammographic masses by deep learning
Xiaoyong Zhang, Takuya Sasaki, Shintaro Suzuki, Yumi Takane, Yusuki Kawasumi, Tadashi Ishibashiz, Noriyasu Homma, Makoto Yoshizawa · 2017
Classification of benign and malignant masses in mammograms is one of the most difficult tasks in development of mammographic computer-aided diagnosis (CAD) system. This paper presents a deep learning-based method that utilizes a deep convolutional neural network (DCNN) to classify mammographic masses into two classes: benign and malignant masses. In order to train the DCNN for mass classification, a transfer learning strategy which pre-trains the DCNN on a large-scale natural image database and subsequently fine-tunes the DCNN on a relative small-scale mam-mogram database is used in this study. We test our method on the mammogram database and evaluate the classification performance using a receiver operating characteristic (ROC) curve. The experimental results demonstrate that the area under the curve (AUC) of ROC is about 0.8 that is closed to the performance of radiologist.