Developing a personalized breast cancer screening tool using sequential mammograms

项目摘要 目前的乳腺癌筛查建议基本上是一种一刀切的方法， 因此在有效性和资源利用方面不是最佳的。这是因为典型的方法 重点是寻找那些患乳腺癌的风险“高于平均水平”的女性亚组， 积极推广其他成像技术。然而，大多数女性（约70%） 乳腺癌没有任何已知的危险因素。此外，大多数妇女（约88%） 永远不会患乳腺癌，这些女性从乳腺癌筛查中受益最少。最大化 为了使所有妇女受益并将可能的危害降至最低，研究人员提倡使用 女性个体患乳腺癌的风险。要做到这一点，必须有一个标记，可以提供准确的 近期乳房X线摄影可检测的乳腺癌（mBCa）风险，以确定女性非常高或非常低 近期mBCa风险。本申请的目的是提供以人为中心的mBCa风险标志物，因此， 提供个性化的筛选策略。我们假设我们可以利用时间变化和横向 通过一种新的成像变换从连续乳房X线照片中提取的图像差异， 基于图像的风险标志物，可以为女性提供准确的近期mBCa风险， 我们将建立一个数据库（N= 1，200，400例病例和800例对照）， 从40岁以上女性的病历中收集的（≥ 5年）全视野数字乳腺X线照片， 开发和额外的独立验证数据集（N = 600，200例病例，400例对照）进行验证。 我们将使用新的氡累积分布变换（RCDT）开发特定年份的风险标记， 卷积神经网络（CNN）和传统的非成像标记（如年龄）。RCDT有效 比较任何两个侧面和颞部乳房X光片，并突出两者之间的差异， 必须明确地对齐两个图像。我们将使用CNN作为强大的成像标记来分析结果 乳房X光片的RCDT图像。使用统计方法处理基于风险的纵向数据 套，我们将结合联合收割机成像为基础的风险标志物和传统的非成像风险因素， 近期风险标志物，一种用于准确预测几年内患有mBCa的极高风险， 另一个是预测几年内mBCa的风险非常低。高风险和低风险的标志物将被 分别进行优化，以最大限度地提高准确预测的高风险和低风险群体的规模。

Project Summary The current breast cancer screening recommendations are essentially a one-size fits all approach and, therefore, not optimal in terms of effectiveness and resource utilization. This is because the typical approach focuses on finding subgroups of women who are at “higher than average risk” for developing breast cancer and aggressively promoting additional imaging techniques. However, most women (approximately 70%) who get breast cancer do not have any known risk factors. In addition, the majority of women (approximately 88%) never get breast cancer and these women benefit the least from breast cancer screening. To maximize the benefit to all women and minimize possible harms, investigators have advocated personalized screening using a woman's individual breast cancer risk. To do so, it is essential to have a marker that can provide an accurate near term mammography-detectable breast cancer (mBCa) risk to identify women with very high or very low near term mBCa risk. The goal of this application is to provide person-centered markers of mBCa risk, thus, offering a personalized screening strategy. We hypothesize that we can use temporal changes and lateral differences in images extracted by a novel imaging transformation from sequential mammograms to develop image-based risk markers that can provide women with an accurate near-term mBCa risk from their last negative mammography exam. We will build a database (N= 1,200, 400 cases and 800 controls) of sequential (≥ 5 years) full field digital mammograms collected from the medical records of women over 40 years of age for development and additional independent validation dataset (N = 600, 200 cases, 400 controls) for validation. We will develop year-specific risk markers using a novel Radon Cumulative Distribution Transform (RCDT), convolutional neural network (CNN), and traditional non-imaging markers (such as age). RCDT effectively compares any two lateral and temporal mammograms and highlights differences between the two without having to explicitly align the two images. We will use CNN as a robust imaging marker to analyze the resulting RCDT images from mammograms. Using a statistical approach for handling longitudinal data based on risk sets, we will combine imaging-based risk markers and conventional non-imaging risk factors to develop two near-term risk markers, one for accurately predicting very high risk of having mBCa within a few years and another for predicting very low risk of having mBCa within a few years. High-risk and low-risk markers will be optimized separately to maximize the sizes of accurately predicted high and low risk groups.