Breast Cancer Risk Assessment via Temporal Changes in Longitudinal Screening Mammograms
DescriptionThe purpose of this study is to implement a deep learning-based computational model to capture the developing asymmetry over four longitudinal negative/benign screening mammogram examinations for predicting breast cancer risk in a case-control setting. We retrospectively identified digital mammograms obtained between 2007 and 2014 at a single institution for cancer and control (were not diagnosed with cancer in this time interval) patients. For each patient, four longitudinal negative/benign prior mammogram examinations were collected(no more than three years between each prior). We designed a deep learning structure (denoted by LRP-Net) to capture the “developing asymmetry”, i.e., the spatiotemporal imaging feature changes of bilateral breast tissue over longitudinal mammogram examinations, to predict the outcome (i.e., cancer vs. control status).
We studied the performance of our proposed method (i.e. LRP-Net) under several different settings, including using a different number of examinations as well as without explicitly capturing the “developing asymmetry” by a “loose” model. All models were evaluated by five-fold cross-validation using the area under the receiver operating characteristic curve (AUC) as the performance metric. The LRP-Net achieved an AUC of 0.67 on using all the four priors, which is significantly higher (p=0.03) than the AUC of 0.64 achieved without explicitly capturing the “developing asymmetry” among the four priors by the “loose” model. This preliminary study shows that the proposed deep learning structure may capture spatiotemporal information related to the developing asymmetry on four longitudinal prior mammogram examinations for breast cancer risk prediction.