Deep learning modeling using normal mammograms for predicting breast cancer risk

• DOI: 10.1002/mp.13886
• Publication Date: 2019-10-31T01:18:43Z
• Authors: Dooman Arefan; Aly A. Mohamed; Wendie A. Berg; Margarita L. Zuley; Jules H. Sumkin; Shandong Wu
• Keywords: Digital Mammography

Purpose To investigate two deep learning‐based modeling schemes for predicting short‐term risk of developing breast cancer using prior normal screening digital mammograms in a case‐control setting. Methods We conducted retrospective Institutional Review Board‐approved study on cohort 226 patients (including 113 women diagnosed with and controls) who underwent general population screening. For each patient, (i.e., negative or benign findings) mammogram examination [including mediolateral oblique (MLO) view craniocaudal (CC) images] was collected. Thus, total 452 images (226 MLO CC images) this were analyzed to predict the outcome, i.e., (cancer cases) remaining cancer‐free (controls) within follow‐up period. implemented an end‐to‐end learning model GoogLeNet‐LDA compared their effects several experimental settings mammographic inputting different subregions models. The proposed models also logistic regression density. Area under receiver operating characteristic curve (AUC) used as performance metric. Results highest AUC 0.73 [95% Confidence Interval (CI): 0.68–0.78; view] when whole‐breast 0.72 (95% CI: 0.67–0.76; + view) dense tissue, respectively, input. GoogleNet‐LDA significantly (all P < 0.05) outperformed GoogLeNet all experiments. consistently more predictive than both models, regardless input subregions. Both exhibited superior percent density (AUC = 0.54; 95% 0.49–0.59). Conclusions approach can images. Larger studies are needed further reveal promise enhancing imaging‐based assessment.