Convolutional Neural Networks and Feature-Visualization for Pathology Classification in Mammograms

Breast cancer is the second deadliest cause of cancer in women. Mammography allows early detection and diagnosis of pathologies, allowing the radiologist to investigate the presence of anomalies, both different types of mass and calcifications. Convolutional neural networks (CNNs) have been successfully applied on medical images for multiple problems. Once trained, after proper analysis and interpretation, the information contained in CNNs can be useful to expand the knowledge about the problem under study. This paper addresses the pathology classification in mammograms with CNNs and their subsequent analysis by means of feature-visualization. It is based on the CBIS-DDSM dataset, considering transfer-learning and networks trained from scratch. The accuracy achieved for mass vs calcification cases was 0.911 +/- 0.014, exceeding that reported in the state of the art for transfer-learning and achieving the performance of networks specifically designed for mammography. The feature-visualization allowed to achieve relevant conclusions about the features extracted, proving being a suitable starting point for the analysis of medical images.