Application of DenseNets for Classification of Breast Cancer Mammograms

Abstract

In this study, we focus on the problem of a breast cancer diagnosis using mammography images by classifying them as belonging either to a negative or to a malignant mass class. We explore the potential of densely connected convolutional neural network (DenseNet) architectures by comparing its three different variants that were trained to classify the abnormalities in breast tissue. The models have been tested in a series of systematic experiments. With a limited dataset (2247 images per class), it was necessary to perform tests to verify whether the amount of data used in this work is sufficient to allow for the conclusion that the experimental results are not dependent on the subset of the data. The training was conducted using stratified 10-fold cross-validation to obtain statistically reliable metrics estimates. DenseNet-201 was found to be the best model achieving: 0.96 value for area under the curve (AUC), 0.92 for precision, 0.90 for recall, and 91% for accuracy.