Lung Imaging based Risk Score (LunIRIS): Decision support tool for screening CT

摘要：国家肺部筛查试验(NLST)的最新数据表明，每年的低剂量胸部 对吸烟患者进行CT扫描，可以及早发现肺癌，提高存活率。合作医疗/医疗保险 因此批准了CT扫描用于肺癌筛查，并批准了退伍军人事务部国家健康中心 促进和疾病预防也采用了类似的方法。退伍军人(退伍军人)人数正在增加 与普通人群相比，由于吸烟率较高和 在服兵役期间接触其他致癌物质的可能性增加。退伍军人管理局关心的是 每年有670万名退伍军人，其中许多人有很长的吸烟史，其中大部分是年长的男性退伍军人。在最近的一项研究中， 来自美国八个退伍军人中心的调查人员在两年内对2000多名退伍军人进行了筛选 来自NLST的标准。在接受筛查的2106名退伍军人中，共有1257人(59.7%)有结节，其中 需要跟踪的1184例(56.2%)。几乎所有的阳性结果都是癌症阴性，产生了一个错误的- 以人为本解释的阳性率为97.5%。在普通人群中，许多肺结节 在胸部CT上被读者识别为“不确定的”或“可疑的”会引发额外的手术 介入治疗(约5000美元-25K美元/患者)和CT检查，但30%的结节在随后的活检或 切除被认为是良性的。目前筛查结节诊断假阳性率较低 CT检查会导致患者焦虑，这也是肺癌筛查依从性差的原因之一。作为一名 结果，迫切需要更好的基于图像的决策支持工具来改进肺癌筛查。 PiAnant Madabhushi和他的团队开发了新型的计算机化图像分析和图案 改进常规筛查中区分癌性和非癌性结节的识别工具 胸部CT扫描。一项重大突破是开发了一种新型的成像标记物，称为“血管 曲度“用于定量表征肺结节血管构筑的复杂性 胸部CT扫描：良性和恶性血管弯曲的测量有显著不同 肺结节。此外，我们的团队还确定了其他高度预测的图像功能，旨在 捕捉(1)结节内和紧邻结节外的微结构的细微纹理图案，以及(2) 结节微妙的3D形状图案。这些成像标记中的每一个都被独立地显示为 在区分接收器工作特性曲线(AUC)下的面积为77%-87% 在N=145名患者的验证集中，恶性结节与良性结节之间的关系。相比之下，在这群人中，有一个专家胸膛 放射科医生和肺科医生的AUC最高为69-72%。更有趣的是，在这个队列组合中 基于机器的与人类读者的解释导致AUC值提高了30% 人类读者。 在我们目前令人印象深刻的结果的基础上，在这项研究中，我们建议继续优化我们的 计算机化决策支持技术(基于肺部成像的风险评分(LUNIRiS))，将风险评分指定为 胸部CT扫描显示癌变为结节。在目标1中，我们将确定内部和周围的最佳组合 构建LUNIRiS的结节结构、3D形状、边缘锐度和血管弯曲度测量 通过使用超过N=300名患者的队列进行软件编程。在AIM 2中，LUNIRiS将独立 在来自克利夫兰退伍军人管理局的N=300个回顾病例上进行了验证。然后，我们将在 克利夫兰退伍军人事务部在目标3中定量评估其作为决策支持工具的作用。在一组独立的 来自克利夫兰退伍军人管理局的资深患者、放射科医生和肺科医生的250例CT筛查检查 首先独立阅读扫描；经过一段时间后，他们将执行第二次解释 月亮星。然后将使用和不使用LUNIRiS的解释结果进行比较，以评估 月亮星。

ABSTRACT: Recent data from the National Lung Screening Trial (NLST) suggest that annual low-dose chest CT scans in patients who smoke, leads to early detection of lung cancer and improves survival. CMS/Medicare has consequently approved CT scans for lung cancer screening, and the VA National Center for Health Promotion and Disease Prevention has adopted a similar approach. The Veteran (VA) population is at increased risk of developing lung cancer as compared to the general population because of higher smoking rates and increased likelihood of exposure to other carcinogens during their military service. The VA system cares for some 6.7 million mostly older male veterans each year, many of whom have long smoking histories. In a recent study, investigators from eight VA centers across the U.S. screened more than 2,000 Veterans over two years using criteria from the NLST. Among the 2,106 Veterans screened, a total of 1,257 (59.7%) had nodules, of which 1,184 (56.2%) required tracking. Nearly all of the positive results were negative for cancer, producing a false- positive rate of 97.5% for human-based interpretation. In the general population, many of the lung nodules identified by human readers as “indeterminate” or “suspicious” on chest CT trigger additional surgical interventions (~$5K-$25K/patient) and CT exams, but >30% of these nodules on subsequent biopsies or resection are identified as being benign. The current low false positive rate in diagnosis of nodules on screening CT exams results in patient anxiety, and one of the reasons for poor compliance in lung cancer screening. As a result, there is an urgent need for better image based decision support tools for improving lung cancer screening. PI Anant Madabhushi and his team have developed novel computerized image analysis and pattern recognition tools for improved discrimination of cancerous from non-cancerous nodules on routine screening chest CT scans. A significant breakthrough has been in developing a novel imaging marker called “vessel tortuosity” for quantitatively characterizing the architectural complexity of the vasculature of a lung nodule on chest CT scans; measurements of vessel tortuosity being significantly different between benign and malignant lung nodules. Additionally our group has also identified other highly predictive image features that aim to capture (1) subtle textural patterns of the microarchitecture within and immediately outside the nodule, and (2) subtle 3D shape patterns of the nodule. Each of these imaging markers has been independently shown to have an area under the receiver operating characteristic curve (AUC) ranging from 77-87% in distinguishing malignant from benign nodules in a validation set of N=145 patients. By contrast, on this cohort an expert chest radiologist and pulmonologist had a maximum AUCs of 69-72%. More interestingly, on this cohort combining machine based interpretations with human readers resulted in an improvement of 30% in the AUC value for the human readers. Building on our current impressive results, in this study we propose to continue to optimize our computerized decision support technology (Lung Imaging based Risk Score (LunIRiS)) to assign a risk score of malignancy to a nodule on a chest CT scan. In Aim 1 we will identify the best combination of intra- and peri- nodule texture, 3D shape, margin sharpness and vessel tortuosity measurements for constructing the LunIRiS software program by employing a cohort of over N=300 patients. In Aim 2, LunIRiS will be independently validated on N=300 retrospective cases from the Cleveland VA. We will then deploy the LunIRiS program at the Cleveland VA in Aim 3 to quantitatively evaluate its role as a decision support tool. On an independent cohort of N=250 CT screening exams from Veteran patients, radiologists and pulmonologists at the Cleveland VA will first independently read the scans; following a wash out period they will perform a second interpretation with LunIRiS. Interpretation results with and without LunIRiS will then be compared to evaluate additional benefit of LunIRiS.