Artificial Intelligence for Improved Breast Cancer Screening Accuracy: External Validation, Refinement, and Clinical Translation

项目总结 筛查乳房X光检查挽救了生命，但仅靠人类解释是不完美的，而且与 重大危害包括每年约30,000例漏诊乳腺癌和约380万次假阳性考试 仅在美国。传统的计算机辅助检测未能提高筛查精度，部分原因是 几十年前在小数据集上训练和测试的软件的静态性质。改良技术的最新进展 计算机处理能力、基于云的数据存储能力以及大型成像数据集的可用性 这让人们再次对将人工智能(AI)应用于乳房X光检查的解释感到兴奋。 我们提出了一种独特的学术-行业合作伙伴关系，以验证、改进、扩展和临床转换我们的 成熟的2D乳房X光摄影人工智能算法用于3D乳房X光摄影解释。我们的团队帮助组织和 引领逆向工程评估和方法(DREAM)数字乳房摄影对话 Challenger，一个开放的众包AI算法挑战，提供了640,000个数字2D乳房X光检查 向全球1,200个编码团队提供图像和相关的临床元数据。我们的行业合作伙伴 格兰特，DeepHealth，Inc.，是梦想挑战赛中表现最好的团队。拥有美国50%的工厂 现在提供3D乳房X光检查，成像数据增加了50到100倍，这代表着一种新的 放射科医生和人工智能算法开发人员的关键障碍。到目前为止，几乎没有什么出版物 解决人工智能增强的3D乳房X光检查解释，这是新兴的筛查选择。 我们将验证我们的梦后2D乳房X光检查算法(目前该算法可与人类解释相媲美 使用加州大学洛杉矶分校的雅典娜乳房健康网络，这是最大的基于人群的乳房成像之一 注册处。我们将在放射科专家的监督下增强我们的2D AI算法，并检查 添加新的非成像数据参数，包括基因突变和肿瘤分子亚型数据 帮助训练人工智能模型，以识别更多具有临床意义的癌症。我们将使用几种新的技术 从2D到3D乳房X光检查的算法方法，在我们的初步研究中显示 提高了准确性，而不仅仅是放射科医生的解释。最后，我们将进行一系列的解读 研究以确定“黑匣子”输出和放射科医生翻译之间的最佳接口， 这仍然是一个未被充分研究的话题。每年有4000万美国女性接受筛查，似乎 总体准确度的小幅改善仍将意味着以人口为基础的结果的显著改善。 总而言之，我们组建了一支在机器/深度领域拥有专业知识的无与伦比的多学科团队 学习，乳腺癌筛查准确性，医学，肿瘤学，放射学，成像技术评估， 和生物统计学。我们拥有久经考验的强大协作记录，并处于有利地位来验证、 增强、扩展和转换我们久经考验的2D AI算法，以提高3D乳房X光摄影的精度。我们的新产品 最终用户工具将有助于将常规筛查的天平推向更大的好处而不是伤害。

PROJECT SUMMARY Screening mammography saves lives but human interpretation alone is imperfect and is associated with significant harms including ~30,000 missed breast cancers and ~3.8 million false-positives exams each year in the U.S. alone. Traditional computer-aided detection failed to improve screening accuracy, in part due to the static nature of software trained and tested on small datasets decades ago. Recent advances in improved computer processing power, cloud-based data storage capabilities, and availability of large imaging datasets have led to renewed excitement for applying artificial intelligence (AI) to mammography interpretation. We propose a unique academic-industry partnership to validate, refine, scale, and clinically translate our proven 2D mammography AI algorithm to 3D mammography interpretation. Our team helped organize and lead the Dialogue for Reverse Engineering Assessments and Methods (DREAM) Digital Mammography Challenge, an open crowdsourced AI algorithmic challenge that provided >640,000 digital 2D mammogram images and associated clinical metadata to >1,200 coding teams worldwide. Our industry partner for this grant, DeepHealth, Inc., was the top performing team in the DREAM Challenge. With >50% of U.S. facilities now offering 3D mammography for screening, the 50-to-100-fold increase in imaging data represents a new critical barrier for both radiologists and AI algorithm developers. To date, there have been few publications addressing AI-enhanced interpretation of 3D mammography, the emerging screening exam of choice. We will validate our post-DREAM algorithm for 2D mammography (which currently rivals human interpretation alone) using UCLA's Athena Breast Health Network, one of the largest population-based breast imaging registries. We will enhance our 2D AI algorithm with expert radiologist supervision and examine the impact of adding novel non-imaging data parameters, including genetic mutation and tumor molecular subtype data, to help train the AI model to identify more clinically significant cancers. We will use several novel technical algorithmic approaches to scale from 2D to 3D mammography which, in our preliminary studies, have shown improved accuracy beyond radiologist interpretation alone. Finally, we will perform a series of interpretive studies to identify the optimal interface between “black box” outputs and radiologist interpreters, which remains an understudied topic. With >40 million U.S. women undergoing screening each year, seemingly small improvements in overall accuracy would still imply significantly improved population-based outcomes. In summary, we have assembled an unparalleled multidisciplinary team with expertise in machine/deep learning, breast cancer screening accuracy, medicine, oncology, radiology, imaging technology assessment, and biostatistics. We have a proven track record of strong collaboration and are well positioned to validate, enhance, scale, and translate our proven 2D AI algorithm for improved 3D mammography accuracy. Our new end user tool will help tip the balance of routine screening towards greater benefits than harms.