Machine learning for risk-adjusted breast MRI screening

总结 磁共振成像（MRI）是迄今为止诊断乳腺癌最敏感的成像方式。 具有强烈家族史或相关基因突变的女性患乳腺癌的风险增加， 建议每年参加MRI筛查。然而，在这一高风险队列中的检出率是 小，促使人们希望减少不必要的MRI检查。该项目的基本假设是， 筛查组群，可以基于乳房的外观来估计未来癌症的个体风险， 核磁共振和乳房X光检查。在初步工作中，我们已经确定了低风险的妇女， 在没有遗漏新癌症的情况下省略了一次筛查。这个低风险亚组的发现是 通过前期工作中开发的现代深度学习工具实现。Memorial Sloan Kettering癌症 中心（MSK）已经积累了一个数据库，其中包含沿着18年来约70，000例乳腺MRI检查， 患者的临床结果。这种前所未有的资源使现代机器学习的训练成为可能 “从地面向上”，以提取和分类体积MRI特征。该项目的具体目标如下： 如下目标1（数据策展）：对MSK积累的大型数据集进行系统分析需要仔细 策展，包括图像内容、图像质量、病理结果、临床随访以及人口统计学 和基因组信息。这个目标的结果是一个精心策划的数据集，可以广泛地受益于未来的技术 乳腺诊断的努力。目标2（深度学习）：为了量化风险分层，我们建议 使用经过训练的现代深度网络分析MRI扫描，以识别位置和范围 一种癌症。然后，我们将转移这些训练过的网络以及在 乳房X光检查的诊断和风险评估的任务。本目标的预期结果具有预测性 在诊断和分割方面具有人类水平性能的模型。目标3（风险调整筛选）： 减少筛查负担，同时保持敏感性，我们将估计发现恶性肿瘤的风险。 根据目前的MRI检查和最近的乳房X光检查以及患者 信息.机器估计的风险将用于回顾性分析，以确定主要 结果，即通过安排更长的筛选间隔可以省略的检查次数 而不影响灵敏度。这将在MSK、杜克和约翰斯的新累积数据上重复 霍普金斯大学（JHU）作为二级研究中心。一旦得到验证，风险预测模型将被公开 以鼓励数据共享和临床采用。在过去两年中进行的初步工作 多年来，汇集了一个独特的跨学科团队，包括乳腺MRI的临床研究人员， MSK以及CCNY、杜克和JHU的机器学习和医学成像专家。平台技术 这将在这里开发的适用于乳腺癌以外，和迁移学习的方法适用 特别是对于具有更有限数据集的癌症。

SUMMARY Magnetic Resonance Imaging (MRI) is the most sensitive imaging modality for breast cancer diagnosis to date. Women with a strong family history or related genetic mutations have an elevated risk of breast cancer and are recommended to participate in yearly MRI screenings. However, the rate of detection in this high-risk cohort is small, prompting a desire to reduce unnecessary MRI exams. The basic hypothesis of this project is that within the screening cohort the individual risk of a future cancer can be estimated based on the appearance of breast MRI and mammograms today. In preliminary work we have already identified low-risk women that could have omitted a screening session without missing a new cancer. The discovery of this lower-risk subgroup was made possible by modern deep-learning tools developed in preliminary work. Memorial Sloan Kettering Cancer Center (MSK) has accrued a database of approximately 70,000 breast MRI exams over 18 years along with the patients’ clinical outcomes. This unprecedented resource enables the training of modern machine learning “from the ground-up” to extract and classify volumetric MRI features. The specific aims of this project are as follows. Aim 1 (Data curation): Systematic analysis of the large dataset accrued at MSK requires careful curation including image content, image quality, pathology results, clinical follow-up, as well as demographic and genomic information. The outcome of this Aim is a curated dataset that can broadly benefit future technical efforts in breast diagnosis. Aim 2 (Deep learning): To make risk stratification quantitative we propose to analyze the MRI scans using modern deep networks that have been trained to identify the location and extent of a cancer. We will then transfer the MRI features of these trained networks as well as networks trained on mammograms to the task of diagnosis and risk assessment. The intended outcome of this Aim are predictive models with human-level performance at diagnosis and segmentation. Aim 3 (Risk adjusted screening): To reduce the burden of screening while maintaining sensitivity we will estimate the risk of finding a malignant tumor in the future, based on the present MRI exam and most recent mammogram as well as patient information. The machine-estimated risk will be used in a retrospective analysis to determine the primary outcome, namely, the number of exams that could have been omitted by scheduling a longer screening interval without compromising sensitivity. This will be repeated on newly accrued data at MSK, Duke and Johns Hopkins University (JHU) as secondary sites. Once validated, the risk-prediction model will be publicly released to encourage data sharing and clinical adoption. The preliminary work performed over the last two years has brought together a unique interdisciplinary team including clinical investigators on breast MRI at MSK, and machine-learning and medical imaging experts at CCNY, Duke and JHU. The platform technology that will be developed here is applicable beyond breast cancer, and the transfer learning approach applicable in particular to cancers with more limited datasets.