Risk stratifying indeterminate pulmonary nodules with jointly learned features from longitudinal radiologic and clinical big data

项目摘要 不确定性肺结节（IPN）是非常普遍的放射学发现，代表了实质性的肺结节。 这对患者和国家卫生保健系统造成了负担，因为它们提出了诊断挑战。 迫切需要将IPN准确地分层为低和高恶性风险亚组， 与标准化和经过充分验证的临床管理路径相关。临床预测 模型有潜力以可扩展、成本效益高、自动化和非侵入性的方式做到这一点，但 必须在预测准确性方面取得进展，才能在医疗实践中产生实质性影响。 一个未探索的方向在这一领域是整合重复测量的计算机断层扫描（CT）研究 以及在同一预测模型内的临床收集的信息。这种联合学习策略具有优势 潜在的模拟动态放射学变化，如结节生长速率， 临床变量，如吸烟模式和实验室异常。这一观点激发了 假设整合来自纵向成像和纵向临床记录的信息将 改进个性化IPN危险分层和肺癌亚分类从临床医生的透镜来看， 考虑到诊断中涉及许多随时间变化的模式， 决策。该项目利用人工智能（AI）和放射组学方法分析三个 回顾性队列，可能增加大型前瞻性队列。目标1中的拟议工作将 扩展现有的深度学习技术，在纵向图像上训练联合学习模型， 临床记录，以估计联合队列中IPN患者随时间推移的恶性概率 超过2000个主题。这种新的战略将评估对单模态模型和公约 在实践中使用的模型。评估将比较模型在将IPN分层为 低风险和高风险亚组作为临床效用的量度。目标2询问放射组学中的纵向变化是否 特征可以区分惰性和侵袭性肺腺癌，其他肺癌亚型， 和肺转移。这项拟议的研究将是第一个全面描述纵向 肺癌亚型的放射组学，并有可能确定新的纵向放射组学特征 这将有助于对反复发作的患者进行早期IPN评估和非侵入性肺癌亚分类， 显像总而言之，拟议的研究询问，如果巧妙地整合纵向信息， 可以利用不同的模式来推进IPN风险分层和肺癌亚分类。这 奖学金将在范德比尔特大学与导师在高度合作的培训环境中进行 在医学成像AI，肺肿瘤学，生物医学信息学，放射学和生物统计学。拟议 研究和培训计划是协同设计的，以最终准备候选人的医生 科学家的职业生涯在工程创新和精密肿瘤学的交叉点。

PROJECT SUMMARY Indeterminate pulmonary nodules (IPNs) are highly prevalent radiologic findings that represent a substantial burden to patients and the national health care system because of the diagnostic challenge they present. There is a dire need to accurately stratify IPNs into low and high malignancy risk subgroups which are associated with clinical management pathways that are standardized and well validated. Clinical prediction models have the potential to do so in a scalable, cost-efficient, automated, and noninvasive manner, but advances in predictive accuracy must be made before they can make a substantial impact in medical practice. An unexplored direction in this area is integrating repeated measures of computed tomography (CT) studies and clinically-collected information within the same prediction model. This joint learning strategy has advantage of potentially modeling how dynamic radiologic changes like nodule growth rate vary with the trajectory of clinical variables such as smoking patterns and laboratory abnormalities. This perspective motivates the hypothesis that integrating information from longitudinal imaging and longitudinal clinical records will improve personalized IPN risk stratification and lung cancer subclassification From a clinician’s lens, this finding would not be surprising given the many time-varying modalities that are involved in diagnosis and decision making. This project leverages artificial intelligence (AI) and radiomic methods to analyze three retrospective cohorts with the possible addition of a large prospective cohort. The proposed work in Aim 1 will extend upon existing deep learning techniques to train a joint learning model on longitudinal images and clinical records to estimate the malignancy probability across time in patients with IPNs in a combined cohort exceeding 2000 subjects. This novel strategy will be evaluated against single-modality models and convention models that are used in practice. The evaluation will compare the models’ performance in stratifying IPNs into the low and high risk subgroups as a measure of clinical utility. Aim 2 asks if longitudinal change in radiomic features can distinguish between indolent and aggressive lung adenocarcinoma, other lung cancer subtypes, and pulmonary metastases. The proposed study will be the first to comprehensively characterize longitudinal radiomics across lung cancer subtypes and has the potential to identify novel longitudinal radiomic features that will aid early IPN evaluation and noninvasive lung cancer subclassification in patients with repeated imaging. In summary, the proposed research asks if clever integration of longitudinal information across different modalities can be leveraged to advance IPN risk stratification and lung cancer subclassification. This fellowship will be conducted at Vanderbilt University in a highly collaborative training environment with mentors in medical imaging AI, pulmonary oncology, biomedical informatics, radiology, and biostatistics. The proposed research and training plans are synergistically designed to ultimately prepare the candidate for a physician scientist career at the intersection of engineering innovation and precision oncology.