Abstracts

RETURN TO ABSTRACT LISTING


1975. Can a Deep Learning Mammography-Based Model Better Predict Breast Cancer Risk in Women with a Personal History of Breast Cancer?
Authors * Denotes Presenting Author
  1. Shinn-Huey Chou *; Massachusetts General Hospital
  2. Adam Yala; Massachusetts Institute of Technology
  3. Peter Mikhael; Massachusetts Institute of Technology
  4. Leslie Lamb; Massachusetts General Hospital
  5. Regina Barzilay; Massachusetts Institute of Technology
  6. Constance Lehman; Massachusetts General Hospital
Objective:
To assess a mammography-based deep learning (DL) model’s ability to predict second breast cancer (SBC) risk in the large, growing population of women with a personal history of breast cancer (PHBC) and to compare performance with version 8 of the Tyrer-Cuzick model (TC8).

Materials and Methods:
This study included consecutive screening mammograms (MG) from 80,818 women between 1/1/2009 and 12/31/2016. To develop the DL risk assessment model, the training and validation sets were randomly assigned ¬210,819 exams in 56,831 women (3,520 exams in 1,611 women with PHBC) and 25,644 exams in 7,021 women (474 exams in 213 women with PHBC), respectively. To assess the model’s ability to predict 5-year SBC risk, the test set included all women diagnosed with SBC within 3 months to 5 years of the index MG and excluded those without SBC who lacked 5-year imaging follow-up, yielding 166 randomly assigned exams in 85 women with PHBC. We obtained cancer outcomes through linkage to a regional tumor registry. We compared DL model performance to TC8 by comparing areas under the receiver-operating-characteristic curve (AUC).

Results:
Mean age of the total 85 women with PHBC in the test set was 61.1 years (range 41 to 89). The test set consisted of 78.9% (131/166) exams in post-menopausal women, 88.0% (146/166) in White women, and 52.4% (87/166) with low MG density. Of the 166 exams in the test set, there were 58 (34.9%, 95%CI: [0.28, 0.42]) that developed 24 SBCs in 5 years. There were no significant differences in patient age (p=0.61), density (p=0.46), and race (p=0.09) between exams with vs. without SBC. Among the 24 SBCs, 15 (62.5%) were screen-detected, 16 (66.7%) were in the ipsilateral breast, 17 (70.8%) were invasive/microinvasive. Median time from primary to SBC was 5 years (range 2-11); median time from the first index MG to SBC was 39.5 months (range 11-56). The DL model showed an AUC of 0.67 (95%CI: 0.556, 0.789) compared to TC8 with an AUC of 0.5 (95%CI: 0.339, 0.653).

Conclusion:
Currently, no reliable tool exists to predict SBC risk among women with PHBC beyond information gleaned from tumor biology and stage of the primary breast cancer. Our DL model demonstrates improved performance in predicting 5-year SBC risk among women with PHBC compared to TC8. A mammography-based DL model can potentially provide a reliable and accurate 5-year risk-prediction tool for breast cancer recurrence among women with PHBC.