Spatially Resolved Single-Cell Morphometry of Benign Breast Disease Biopsy Images Uncovers Quantitative Cytomorphometric Features Predictive of Subsequent Invasive Breast Cancer Risk

Abstract

Currently, benign breast disease (BBD) pathologic classification and invasive breast cancer (BC) risk assessment are based on qualitative epithelial changes, with limited utility for BC risk stratification for women with lower-risk category BBD (ie, nonproliferative disease [NPD] and proliferative disease without atypia [PDWA]). Here, machine learning-based single-cell morphometry was used to characterize quantitative changes in epithelial nuclear morphology that reflect functional/structural decline (ie, increasing nuclear size, assessed as epithelial nuclear area and nuclear perimeter), altered DNA chromatin content (ie, increasing nuclear chromasia), and increased cellular crowding/proliferation (ie, increasing nuclear contour irregularity). Cytomorphologic changes reflecting chronic stromal inflammation were assessed using stromal cellular density. Data and pathology materials were obtained from a case-control study (n = 972) nested within a cohort of 15,395 women diagnosed with BBD at Kaiser Permanente Northwest (1971-2012). Odds ratios (ORs) and 95% confidence intervals (CIs) for associations of cytomorphometric features with risk of subsequent BC were assessed using multivariable logistic regression. More than 55 million epithelial and 37 million stromal cells were profiled across 972 BBD images. Cytomorphometric features were individually predictive of subsequent BC risk, independently of BBD histologic classification. However, cytomorphometric features of epithelial functional/structural decline were statistically significantly predictive of low-grade but not high-grade BC following PDWA (OR for low-grade BC per 1-SD increase in nuclear area and nuclear perimeter, 2.10; 95% CI, 1.26-3.49, and 2.22; 95% CI, 1.30-3.78, respectively), whereas stromal inflammation was predictive of high-grade but not low-grade BC following NPD (OR for high-grade BC per 1-SD increase in stromal cellular density, 1.53; 95% CI, 1.13-2.08). Associations of nuclear chromasia and nuclear contour irregularity with subsequent tumor grade were context specific, with both features predicting low-grade BC risk following PDWA (OR per 1-SD, 1.58; 95% CI, 1.06-2.35, and 2.21; 95% CI, 1.25-3.91, for nuclear chromasia and nuclear contour irregularity, respectively) and high-grade BC following NPD (OR per 1-SD, 1.47; 95% CI, 1.11-1.96, and 1.29; 95% CI, 1.00-1.70, for nuclear chromasia and nuclear contour irregularity, respectively). The results indicate that cytomorphometric features on BBD hematoxylin-eosin-stained images might help to refine BC risk estimation and potentially inform BC risk reduction strategies for BBD patients, particularly those currently designated as low risk.

Keywords: artificial intelligence and machine learning; benign breast disease; breast cancer risk; epithelial nuclear characteristics; single-cell cytomorphometry; stromal inflammation and stromal cellular density.