Background: Most breast biopsies are negative for malignancy but may still indicate elevated risk of invasive breast cancer (IBC) due to the presence of benign breast disease (BBD). Cellular senescence plays a complex but poorly understood role in breast cancer development. Deep learning methods offer a novel approach to predict senescence from biopsy images, potentially enhancing our ability to predict IBC risk in women with BBD.
Methods: We conducted a case-control study, nested within a cohort of 15,395 women biopsied for BBD at Kaiser Permanente Northwest between 1971 and 2006. Cases (n=512) were women who developed a subsequent IBC ≥ one year after the BBD biopsy; controls (n=491) did not develop IBC during the same follow-up period. Using H&E-stained biopsy images, we predicted senescence based on deep learning models trained on replicative senescence (RS), ionizing radiation (IR), and various drug treatments. Age-adjusted and multivariable odds ratios (ORs) and 95% confidence intervals (CI) were estimated using unconditional logistic regression.
Results: The RS- and IR-derived senescence scores in adipose tissue and the RS-derived score for epithelial tissue were positively associated with risk of IBC (adipose tissue - RS model: ORq4 vs q1=1.69, 95% CI 1.03-2.77, and IR model: ORq4 vs. q1=1.73, 95%CI 1.06-2.82; epithelial tissue – RS model: ORq4 vs. q1=1.53, 95% CI 1.05-2.22). The results were stronger among postmenopausal women and women with epithelial hyperplasia with/without atypia. There was increased risk of IBC in those with higher senescence scores in both epithelial and adipose tissue compared with those with low senescence scores in both (IR epithelium-IR fat: ORq2-4 vs. q1=2.14, 95% CI 1.30-3.51; IR epithelium-RS fat: ORq2-4 vs. q1= 2.24, 95% CI 1.15-4.35).
Conclusion: Nuclear senescence scores predicted by deep learning models in breast epithelial and adipose tissue were positively associated with the risk of IBC among women with BBD.