Mammogram Classification with Ordered Loss

Abstract

Breast radiologists inspect mammograms with the utmost consideration to capture true cancer cases. Yet, machine learning models are typically designed to perform a binary classification, by joining several severities into one positive class. In such scenarios with mixed gradings, a reliable classifier would make less mistakes between distant severities such as missing a true cancer case and calling it as normal or vise versa. To this end, we suggest a simple yet elegant formulation for training a deep learning model with ordered loss, by increasingly weighting the loss of more severe cases, to enforce importance of certain errors over others. Training with the ordered loss yields fewer severe errors and can decrease the chances of missing true cancers. We evaluated our method on mammogram classification, using a weakly supervised deep learning method. Our data set included over 16 K mammograms, with a large set of nearly 2,500 biopsy proven cancer cases. Evaluation of our proposed loss function showed a reduction in severe errors of missing true cancers, while preserving overall classification performance in the original task.