Breast cancer: machine learning aimed at triaging for screening, diagnostic workflow, and treatment

Abstract

Traditional biomarkers of breast cancer are dependent on invasive sampling of the areas suspicious of malignancy. In contrast, MRI not only investigates the whole tumor in vivo, but also evaluates the surrounding tissues (e.g., background parenchymal enhancement). The aim of this thesis was to investigate machine learning to support triaging decisions in screening, diagnostic workflow, and treatment by extracting imaging biomarkers from fibroglandular tissue and breast lesions on MRI. We identified quantitative properties of the breast parenchyma on baseline MRI in a screening population of women with extremely dense breasts, and found that breast cancer occurrence in women with higher volumes of enhancing parenchyma was around two times larger than that in women with low volumes of enhancing parenchyma (hazard ratio, 2.09, P = .005). In a population of women with early breast cancer, we constructed a computer-aided diagnosis model and identified malignancy with near-perfect specificity in approximately half of preoperative patients originally indicated for a breast biopsy. Finally, we investigated the imaging phenotype of synchronous cancers, and found that the MRI phenotype of ER+/HER2– breast cancer was different from that of its ipsilateral counterpart; a large MRI phenotype difference was associated with worse prognosis, which may have potential to serve as a noninvasive indicator of long-term prognosis before treatment.