Traditional and Deep Learning Based Methods for Mammographic Image Analysis

Deep learning is becoming more and more mature in the field of machine learning, and therefore the application of deep learning in mammographic image analysis is considerably increasing. Deep learning is a multi-layer neural network based on big data, and the algorithms of deep learning can be divided into four categories: the model based on restricted Boltzmann machine (RBM), convolutional neural network (CNN), the model based on autoencoder (AE), and the model based on sparse coding (SC). In this paper, the development status and existing problems of traditional mammographic image analysis methods, covering abnormality detection, regions of interest (ROIs) segmentation, abnormality classification and computer-aided diagnosis (CAD), are first briefly reviewed. Then deep learning based methods for mammographic image segmentation, classification and CAD are reviewed and compared with the traditional methods in order to explore the advantages and disadvantages. Finally, considering the problems of deep learning in mammographic image analysis, and the limitations of the existing deep learning based methods, its development prospect is discussed. It is concluded that deep learning has great potential, and the related applications in computer-aided diagnosis will be more extensive.