Radiomic and genomic predictors of breast cancer risk

摘要 每年将有超过4万名美国女性死于乳腺癌。筛查乳房X光检查挽救了生命，但也 会带来潜在的危害。针对女性个体风险量身定做的个性化筛查方案 通过对高危妇女进行更密集的治疗，改善致命癌症的早期发现，并减少 对低风险妇女的过度筛查和过度治疗。然而，目前临床上乳腺癌的风险 预测模型在区分高风险和低风险女性方面不够准确。新放射学 深度学习算法，自动从女性筛查中挖掘乳房组织特征宝库 乳房X光检查预测她未来的癌症风险，有巨大的潜力改变乳腺癌筛查， 但尚未得到独立验证。新的乳腺癌多基因风险评分(PR)也显示 承诺改进风险预测，尽管在人口规模上实施仍然代价高昂。我们建议 检查添加放射和基因组风险评分是否可以显著改善当前的临床风险 在Kaiser Permanente‘s登记的17.8万名女性中，基于人群的大型、多样化的预测模型 基因、环境与健康研究计划(RPGEH)的2D全场数字化筛查 乳房X光照相(FFDM)。我们还建议将性能最好的放射学深度学习算法扩展到 在加利福尼亚州和纽约的两个大型医疗保健机构中使用的多样化筛查乳房X光照相系统， 包括在西奈山接受3D数字乳房断层合成(DBT)筛查的50K名妇女 卫生系统(MSHS)。具体目标是：(1)评估放射学深度学习的绩效 乳腺癌风险预测模型，估计它们与5年和10年乳腺癌风险的关系， 并确定关联与已知临床风险因素的独立程度；(2) 确定放射和基因组风险评分是否独立预测乳腺癌风险，并探索 种族/民族和其他临床危险因素的潜在差异；以及(3)转移最佳放射深度 从二维有限差分到三维层析合成的学习算法(S)。拟议中的研究将填补必要的知识 通过验证新的放射组学算法来实现放射组学和基因组学的潜力所需的差距， 量化传统风险因素模型和新多基因模型在模型性能方面的改进 风险评分，探索种族/民族的差异，并将最佳放射学工具扩展到不同的 乳房X光照相系统在两个大型多种族医疗保健环境中使用。 1

ABSTRACT Over 40,000 U.S. women will die of breast cancer each year. Screening mammography saves lives but also results in potential harms. Personalized screening regimens tailored to a woman's individual risk can both improve early detection of lethal cancers through more intensive regimens for high-risk women, and reduce over-screening and over-treatment of low-risk women. However, the current clinical breast cancer risk prediction models are insufficiently accurate for discriminating high-risk and low-risk women. New radiomic deep learning algorithms, which automatically mine troves of breast tissue features from a woman's screening mammogram to predict her future cancer risk, have enormous potential to transform breast cancer screening, but have not been independently validated. New polygenic risk scores (PRS) for breast cancer also show promise for improving risk prediction, although still costly to implement on a population scale. We propose to examine whether adding radiomic and genomic risk scores can significantly improve current clinical risk prediction models in a large, diverse population-based cohort of 178K women enrolled in Kaiser Permanente's Research Program on Genes, Environment and Health (RPGEH) who were screened with 2D full-field digital mammography (FFDM). We also propose to extend the best performing radiomic deep learning algorithms to diverse screening mammography systems utilized in two large health care settings in California and New York, including a cohort of 50K women screened with 3D digital breast tomosynthesis (DBT) in the Mount Sinai Health System (MSHS). The specific aims are to: (1) Evaluate the performance of radiomic deep learning breast cancer risk prediction models, estimate their associations with 5-year and 10-year breast cancer risk, and determine the extent to which the associations are independent of known clinical risk factors; (2) Determine whether radiomic and genomic risk scores independently predict breast cancer risk, and explore potential differences by race/ethnicity and other clinical risk factors; and (3) Transfer the best radiomic deep learning algorithm(s) from 2D FFDM to 3D tomosynthesis. The proposed research will fill essential knowledge gaps needed to realize the potential of radiomics and genomics by validating new radiomic algorithms, quantifying the improvements in model performance above traditional risk factor models and new polygenic risk scores, exploring differences by race/ethnicity, and extending the best radiomic tools to diverse mammography systems utilized in two large multi-ethnic health care settings. 1