Improving breast cancer risk assessment with image-based deep learning models

Abstract

Discriminative models for breast cancer risk prediction are needed in order to provide personalized patient care. Existing breast cancer risk models incorporate information about breast tissue using imaging biomarkers such as density scores. However, these imaging biomarkers are limited in that they suffer from variability in radiologists' assessments and they reduce the rich information contained in the image down to a single number. In this thesis, I present deep learning models that predict breast cancer risk directly from full images, specifically breast MRIs and mammograms. Our image-based deep learning models out-perform existing breast cancer risk models and our own risk-factor-only models. These results demonstrate that full images contain subtle but significant indicators of risk not captured by traditional risk factors, and that deep learning models can learn these patterns directly from the data.