Using artificial intelligence to support efficient same-day diagnostic imaging in breast cancer screening

乳腺癌是美国最常见的实体癌症，也是美国癌症死亡的第二大原因。 女人。多项研究表明，筛查乳房X光检查可以降低乳腺癌相关死亡率， 但是，实施筛查具有低效和局限性，可能会造成潜在的危害。错误的- 阳性结果和等待进一步评估的时间都是有充分证据的乳房X光检查危害。大致 10%的筛查检查被召回进行诊断检查，其中95%被发现为假阳性。 可能导致良性活检和过度诊断。再次被召回的女性比例 根据放射科医生的不同，诊断体检的比例在8%-14%之间。此外，一个人经历的焦虑 最近有异常乳房X光检查的女性意义重大。许多女性睡眠不佳，难以集中注意力，因为 他们正在等待更明确的诊断结果。考虑到筛查的这些局限性，我们的首要目标是 最大限度地减少与召回相关的实践差异，降低患者焦虑，提高患者满意度。 我们建议从以下方面评估引入人工智能(AI)解决方案的可行性和效果 关注(1)降低总体回调率，(2)通过提供即时筛查来提高患者满意度 结果，对于需要进一步诊断检查的妇女，(3)消除筛查和检查之间的延迟 诊断性检查。人工智能解决方案能够以较高的速度即时在线解释筛选考试 体量乳房筛查计划。对于乳房X光检查异常的女性，实时解读 筛查检查允许妇女被安排在同一天进行诊断检查。这个目标是 实现三个目标。AIM 1将验证并集成人工智能算法来对乳房X光照片进行分类筛查 在我们机构的乳房筛查人群中。我们将确保算法以预期的速度运行 水平(即，不低于现有放射科医生的表现)，并集成和改进算法以进行沟通 向目标用户提供清晰、高效的结果。AIM 2将设计和评估当天启用AI的工作流 诊断性检查。我们将分析目前的护理状况，确定实施该计划的障碍， 并对护理途径进行改变，以允许人工智能干预。在目标3中，我们将实施和评估 启用人工智能的当天诊断成像范例的影响分为三个阶段：(1)试点阶段，涉及 在单一地点接受2D筛查乳房X光检查的妇女子组；(2) 实施阶段，涉及更多的妇女在一年中接受2D和3D筛查乳房X光检查 单一成像中心；以及(3)扩展阶段，包括在第二成像中心对妇女进行筛查。 加州大学洛杉矶分校健康中心是一个独特的环境来评估这一范例，因为有大量的筛查检查 每年进行(40,000次检查)及其乳房筛查计划的分布性质 地理上分散的成像中心。该项目的预期结果是一种可概括的方法 评估和整合人工智能算法，以改善护理提供。

Breast cancer is the most common solid cancer and the second leading cause of cancer death among U.S. women. Multiple studies have shown that screening mammography decreases breast cancer-related mortality, but the implementation of screening has inefficiencies and limitations that contribute to potential harms. False- positive results and wait times for further evaluation are well-documented mammographic harms. Approximately 10% of all screening exams are recalled for diagnostic workup, of which 95% are found to be false positives, potentially resulting in benign biopsies and overdiagnosis. The percentage of women recalled for further diagnostic workup varies between 8-14%, depending on the radiologist. Moreover, the anxiety experienced by a woman with a recent abnormal mammogram is significant. Many women sleep poorly and struggle to focus as they await a more definitive diagnostic workup. Given these limitations of screening, our overarching goal is to minimize practice variabilities associated with recalls, reduce patient anxiety, and increase patient satisfaction. We propose to assess the feasibility and effect of introducing artificial intelligence (AI) solution at the point of care to (1) reduce the overall callback rate, (2) increase patient satisfaction by providing immediate screening results, and for women who require further diagnostic workup, (3) eliminate the delay between screening and diagnostic workup. The AI solution enables immediate “online” interpretation of screening exams in a high- volume breast screening program. For women with abnormal mammograms, real-time interpretation of the screening exam permits women to be scheduled for a diagnostic exam on the same day. This goal is accomplished in three aims. Aim 1 will validate and integrate an AI algorithm to triage screening mammograms within our institution’s breast screening population. We will ensure that the algorithm performs at an expected level (i.e., non-inferior to existing radiologist performance) and integrate and refine the algorithm to communicate results clearly and efficiently to target users. Aim 2 will design and assess an AI-enabled workflow for same-day diagnostic exams. We will analyze the current state of care, identify impediments to implementing this program, and develop changes to the care pathway to allow an AI intervention. In Aim 3, we will implement and evaluate the impacts of an AI-enabled same-day diagnostic imaging paradigm in three stages: (1) a pilot stage, involving a subset of women undergoing screening using 2D screening mammography at a single site; (2) an implementation stage, involving a larger group of women undergoing 2D and 3D screening mammography at a single imaging center; and (3) an expansion stage, involving women being screened at a second imaging center. UCLA Health is a unique environment to evaluate this paradigm given the large number of screening exams performed annually (>40,000 exams) and the distributed nature of its breast screening program across twelve geographically separated imaging centers. The expected outcome of this project is a generalizable approach for evaluating and integrating AI algorithms to effect improvements in care delivery.