Fine-Tuning ResNet for Breast Cancer Classification from Mammography

Breast cancer classification from mammography is significant for treatment decisions and assessments of prognosis. However, the traditional classification method is not efficient due to the need for professional domain knowledge, time-consuming, and difficult in extracting high-quality features. Therefore, this paper proposed an automatic classification method based on convolutional neural network (CNN). In this paper, the fine-tuning residual network (ResNet) has been introduced to have good performance, reduce training time, and automatically extract features. Then, a data augmentation policy was adopted to expand training data which can reduce the probability of overfitting caused by small training set. The main contribution of this paper is to introduce transfer learning and data augmentation to construct an automatic mammography classification, which has high prediction performance. Experiments were conducted on a public data set CBIS-DDSM which contains 2620 scanned film mammography studies. The proposed method obtains desirable performances on accuracy, specificity, sensitivity, AUC, and loss, corresponding to 93.15, 92.17, 93.83%, 0.95, and 0.15. The proposed method is of good robustness and generalization.