Applying Deep Learning for the Detection of Abnormalities in Mammograms

Abstract

Medical imaging produces massive amounts of data. Computer aided diagnosis (CAD) systems that use traditional machine learning algorithms to derive insights from the data provided in the medical industry struggle to perform at a competent level regarding sensitivity and false positive minimization. This paper looks at some of the current methods used to improve CAD systems in the domain of forming breast cancer diagnosis with mammograms. This paper presents deep learning models that use Convolutional Neural Networks (CNN) to identify abnormalities in mammographic studies that can be used as a tool for the diagnosis of breast cancer. We run two experimental cases on two public mammogram databases, namely MIAS and the DDSM. Firstly, the abnormality severity was classified. Secondly, the combination of abnormality type and its severity were compared in multi-label classification. Two CNN architectures, namely miniature versions of VGGNet and GoogLeNet, were also compared. We were able to achieve a best AUC of 0.85 for the classification of abnormality severity on the DDSM data set and a best Hamming loss of 0.27 on the MIAS data set for the multi-label classification task.