Imaging biomarkers are representations of an in vivo biological state and phenotype. The incorporation of breast density in breast cancer risk models, as well as state-mandated reporting of mammographic breast density to women, underscores the central role of imaging biomarkers in risk assessment. In this dissertation, I evaluate breast imaging biomarkers from breast MRI and mammography in their role of future risk prediction and treatment response. The chapters, ordered chronologically, show the evolution of my research interests from quantitative imaging science within a well-controlled experimental trial (Chapter 1), to a population-based evaluation of qualitative clinically derived imaging assessments in an observational cohort (Chapter 2), to finally combining quantitative imaging science for comparative evaluations through a population-based pragmatic assessment in a large managed health system (Chapter 3).
Chapters 1 and 2 focus on background parenchymal enhancement (BPE), which describes the natural phenomenon observed on breast MRI in which normal breast tissue demonstrates signal enhancement related to uptake of intravenous contrast. Biologically, BPE is believed to represent tissue “activated” by endogenous hormones (primarily estrogen) and is dynamic in appearance over time and distribution within a woman’s breast tissue. Chapter 1 focuses on manually defined quantitative imaging biomarkers in the experimental I-SPY 2 trial, an on-going multicenter prospective randomized clinical trial framework used to monitor treatment response and assess novel investigational neoadjuvant chemotherapy (NAC) agents for breast cancer. Women with advanced HER2- breast cancer have limited treatment options. Breast MRI functional tumor volume (FTV) is used to predict pathologic complete response (pCR) to improve treatment efficacy. In addition to FTV, background parenchymal enhancement (BPE) may predict response and was explored for HER2- patients in the ISPY-2 TRIAL. We found that among women with HER2- cancer, BPE alone demonstrated association with pCR in women with HR+HER2- breast cancer, with similar diagnostic performance to FTV. BPE predictors remained significant in multivariate FTV models, but without added discrimination for pCR prediction. This may be due to small sample size limiting ability to create subtype specific multivariate models.
Chapter 2 extends BPE evaluation through comparative associations of qualitative BPE and mammographic breast density for future risk in a population-based assessment using the Breast Cancer Surveillance Consortium (BCSC), involving 46 radiology facilities that participate in one of six regional BCSC registries. Higher levels of BPE were found to be associated with future invasive breast cancer risk independent of breast density. The combination of both high BPE and high breast density was associated with higher risk than either factor alone. BPE also demonstrates subtype specific associations with less aggressive disease, although the association with aggressive disease was noted at moderate and marked levels.
Finally, Chapter 3 examines whether using computer vision artificial intelligence (AI)–based computer vision algorithms, most of which are trained to extract features from mammograms to detect visible breast cancer, can also predict future risk using a population-based case cohort from the Kaiser Permanente Northern California managed health system. We found that all AI mammography algorithms evaluated had clinically and statistically significantly higher discrimination than the BCSC clinical risk model for interval cancer and 5-year future cancer risk, indicating their usefulness. The combination of BCSC and AI further improves risk prediction above AI alone, and decreases the gap in future risk performance between AI algorithms. Training AI algorithms to predict longer-term outcomes may yield further improvements, but the potential impact on clinical decisions requires further study.
