Automatic mass detection from mammograms with region-based convolutional neural network

Abstract

Automatic detection of breast mass from mammograms is a challenging task. Recently, Convolution Neural Networks (CNNs) have been proposed to address this task. However, the performance of these CNN-based detection methods is still limited due to high complexity of breast tissue and varying shape of masses. An Automatic Mass Detection framework with Region-based CNN (AMDR-CNN) is presented in this paper, aiming to efficiently exploit informative features from mammograms. Under a hierarchical candidate mass region generation method with a full-size mammogram, the mammogram is greatly simplified and high-quality region proposals are generated. Then, a deeper CNN is introduced, which simultaneously predicts object bounds and scores at each position. In contrast to previous works, the deeper CNN learns the effective features of mass as well as helps produce accurate detection results. The experiments are performed on two public datasets, which achieves a better performance than state-of-the-art algorithms.