Machine Learning-Based Identification of Cardiomyopathy Risk in Childhood Cancer Survivors

项目摘要/摘要 治疗相关的心肌病/心力衰竭(CHF)是儿童过早发病的主要原因 癌症幸存者。鉴于蒽环类药物和相关心脏毒性化疗药物的广泛使用，以及在 再加上胸部放射治疗，超过一半的儿童癌症长期幸存者处于 与年龄匹配的普通人群相比，早期充血性心力衰竭的风险显著增加。目前，国家 国际共识指南建议常规使用二维(2D)超声心动图 筛查这一高危人群的CHF早期体征，特别是左心室收缩功能不全和 LV几何体的变化。目前，二维超声心动图代表了全美的护理标准 其广泛的可用性，相对较低的成本，以及避免电离辐射或镇静。不过， 二维超声心动图的局限性包括较大的患者内和观察者间的可变性。因此，当前 基于超声心动图的监测仍然具有有限的敏感性，通常需要进行系列研究。 在患者被确定为有潜在异常之前。尽管没有足够的证据来指导 针对儿科癌症幸存者的CHF治疗：非癌症相关性心肌病的证据 儿童和成人都认为早期干预可以缓解或延缓CHF的进展。因此， 改进儿童癌症幸存者早期心力衰竭的检测方法可能具有重要的临床意义 这意味着什么。深度学习是机器学习的一个子领域，它可以自动地从大量的数据中提取模式 非结构化数据集，如医学图像，并正越来越多地被用于疾病的医学 诊断以及疾病的发病和结果预测。我们建议利用一个独特的成像数据集 我们从儿童肿瘤学小组(COG)聚集起来，该小组是NCI赞助的国家临床中心的一部分 试验网络和社区肿瘤学研究计划，探索DL用于增强的潜力 心力衰竭的检测。我们有100多名儿童癌症幸存者的纵向超声心动图数据 发生了CHF，超过350人没有，他们都是使用标准化标准定义的，代表成像 &gt；3000个个体超声心动图的资料库(还在不断增加)。使用该现有的和临床注释的数据集， 我们建议：1)使用深度卷积神经网络(DCNN)，识别出DL-1的最优工艺。 基于对儿童癌症幸存者充血性心力衰竭的评估；以及2)评估 基于DCNN的心力衰竭诊断前超声心动图对心肌病发病的预测。预期结果 包括DCNN的开发，它将区分异常和正常的超声心动图 患有和不患有充血性心力衰竭的儿科癌症幸存者。优化后，我们将进行初步的 疗效分析以确定存活者提前多少年转变为充血性心力衰竭 一个经过优化的DCNN。

PROJECT SUMMARY / ABSTRACT Treatment-related cardiomyopathy/heart failure (CHF) is a leading cause of premature morbidity in childhood cancer survivors. Given the widespread use of anthracycline and related cardiotoxic chemotherapeutics, and in combination with radiotherapy exposure to the chest, over half of long-term survivors of childhood cancer are at significantly increased risk of early CHF compared with an age-matched general population. Currently, national and international consensus guidelines recommend the routine use of 2-dimensional (2D) echocardiography to screen this high-risk population for early signs of CHF, in particular, left ventricular (LV) systolic dysfunction and changes in LV geometry. At present, 2D echocardiography represents the standard of care across the US given its widespread availability, relatively lower cost, and avoidance of ionizing radiation or sedation. Nevertheless, limitations of 2D echocardiography include greater intra-patient and inter-observer variability. As a result, current echocardiography-based surveillance continues to have limited sensitivity and often requires serial studies before a patient is identified as having a potential abnormality. Although there is insufficient evidence to guide CHF management specific to pediatric cancer survivors, the evidence for non-cancer-related cardiomyopathy in both children and adults suggests that earlier intervention can mitigate or delay CHF progression. Therefore, methods that improve the detection of early CHF in childhood cancer survivors may have important clinical implications. Deep learning (DL), a subfield of machine learning, can automatically extract patterns from large unstructured datasets, such as medical images, and is increasingly being utilized in medicine for disease diagnosis as well as disease onset and outcome prediction. We propose to leverage a unique imaging dataset we have assembled from the Children’s Oncology Group (COG), a part of the NCI-sponsored National Clinical Trials Network and Community Oncology Research Program, to explore the potential of DL for enhanced detection of CHF. We have longitudinal echocardiographic data on over 100 survivors of childhood cancer who developed CHF and over 350 who did not, all defined using standardized criteria, representing an imaging repository of >3000 individual echocardiograms (and growing). Using this extant and clinically annotated dataset, we propose to: 1) Using a deep convolutional neural network (DCNN), identify the optimal process for a DL- based assessment of CHF in pediatric cancer survivors; and 2) Assess the feasibility and preliminary efficacy of DCNN-based prediction of cardiomyopathy onset from pre-CHF diagnosis echocardiograms. Expected results include the development of a DCNN that will differentiate between abnormal and normal echocardiograms from pediatric cancer survivors with and without CHF, respectively. After optimization, we will conduct a preliminary efficacy analysis to determine how many years in advance a survivor's transition to CHF can be predicted using an optimized DCNN.