Identifying Architectural Distortion in Mammogram Images Via a SE-DenseNet Model and Twice Transfer Learning

Breast cancer is a leading threaten of woman's health, and mammography is widely used in the early detection and diagnosis of breast cancer. Mass is a common sign observed on mammograms, and the automatic identification of benign and malignant breast mammography mass images is a challenging task. The CAD techniques have been used to improve the classification performance. Deep learning is powerful classification technique, however, it needs a lot of training data, which is not practical for mass detection or classification. In this paper, based on the observation that natural images are greatly different from medical images, we proposed a new network (SE-DenseNet) which combines the Dense convolutional neural network (DenseNet) and the “Squeeze-and-Excitation” (SE) block, and twice fine-tuning method to classify breast mass. We first use the mass and normal breast to fine-tune our model (trained on ImageNet) and then use breast mass dataset (benign and malignant) to further fine-tuning the model. The correctness of the model is verified by a large number of experiments on the BCDR dataset and quantitative comparison with other methods. The proposed model achieved better performance on the BCDR dataset(AUC is 0.984, accuracy is 0.982).