Novel Measures of Terminal Ductal Lobular Unit Involution for Automated Deep-Learning Assessment of Breast Cancer Risk: A Large-Scale Epidemiological Study

Overview

Harvard Medical School and Beth Israel Deaconess Medical Center Terminal ductal lobular units (TDLUs) are milk-producing glands in the breast. TDLUs shrink (involute) with physiological aging. Re duced TDLU involution is associated with increased breast cancer risk. Traditionally, TDLU involution is assessed manually, but this is both labor-intensive and highly subjective. We established and validated a deep-learning computational method to quantify TDLU involution. We examined data from whole slide images of benign breast disease biopsies from a Nurses’ Health Study (NHS) and NHS II nested case-control dataset (283 cases, 944 controls). We applied our computational method to obtain six TDLU involution measures: TDLU count per unit area, median acini count per TDLU, median TDLU span, median TDLU area, percentage of non-adipose tissue area inside TDLUs, and median acini density. Measures we re placed into quartiles according to the control population. TDLU area percentage was significantly associated with breast cancer risk in Quartile 2 compared to Quartile 1. Median acini density was signifi cantly associated with breast cancer risk in Quartile 2 and Quartile 4. No significant association was observed in other measures. This study paves the road to using our deep-learning TDLU involution method to automate breast cancer risk assessment in other large epidemiological cohorts, a step towards the inclusion of automated TDLU involution measures in clinical breast cancer risk models for the management of high-risk patients.