Improving cardiovascular image-based phenotyping using emerging methods in artificial intelligence

总结/摘要 目标-本提案的目标是开发和优化新型深度学习（DL）辅助方法 提高先天性心脏病（CHD）的诊断和临床决策水平。为实现这些目标将 使用DL、机器学习（ML）和相关方法来提取诊断、生物特征表征， 胎儿超声成像的其他信息。值得注意的是，这项工作包括临床翻译评价 这些方法在跨越二十年，数万名患者的人群范围内的成像收集， 几个临床中心。背景-尽管产前检测CHD有明确的和众多的益处， 胎儿超声在理论上能够检测到超过90%的CHD病变，在实践中，胎儿CHD检测 率接近50%。之前的文献表明，这种惊人的诊断差距的一个关键原因是次优的获取 和胎儿心脏图像的解释。DL是一种新的数据科学技术，在模式方面表现出色 图像识别。DL模型是神经网络架构的设计和调整的函数，并且 用于训练网络的图像数据的管理和处理。初步研究-我们有 组建了一个由超声心动图和冠心病专家组成的多学科团队（Grady、Levine和Arnaout博士）， DL和数据科学（Keiser，Butte和Arnaout博士），统计学和临床研究（Arnaout和 Grady），并获得了数以万计的多中心（UCSF和其他六个中心），多模式胎儿 影像学研究我们已经创建了一个可扩展的图像处理管道，将临床研究转换为图像 数据准备好进行计算。我们设计并训练了DL模型，以在胎儿超声中找到关键的心脏视图， 从这些视图中计算标准和高级胎儿心脏生物统计学，并区分正常 心脏和某些冠心病病变。假设-虽然DL功能强大，但仍需要做大量工作来适应 临床成像，并将其转化为患者人群的临床相关性能。我们假设 一个集成的整体DL/ML方法可以导致胎儿CHD诊断的巨大改善。目的- 为此，本提案的主要目的是（1）开发和优化神经网络架构， 有效的数据输入，以缓解胎儿CHD中DL的关键性能瓶颈;以及（2）部署DL模型 在全人群范围内评估其改善诊断、生物特征和精确度的能力 表型分析超过目前的标准治疗。我们的方法包括DL/ML算法和回顾性成像 分析.环境和影响-这项工作将在一个优秀的研究环境中得到支持 处于数据科学、心血管和胎儿成像以及转化信息学的十字路口。工作 提出将提供有价值的工具和设计和评估数据和算法的见解 为DL在临床相关目标的成像，并将奠定重要的基础DL辅助表型， 临床应用和精准医学研究。

Summary / Abstract Objective — The goal of this proposal is to develop and optimize novel deep learning (DL) assisted approaches to improve diagnosis and clinical decision-making for congenital heart disease (CHD). This will be achieved by using DL, machine learning (ML), and related methods to extract diagnosis, biometric characterizations, and other information from fetal ultrasound imaging. Notably, this work includes a clinical translational evaluation of these methods in a population-wide imaging collection spanning two decades, tens of thousands of patients, and several clinical centers. Background — Despite clear and numerous benefits to prenatal detection of CHD and an ability for fetal ultrasound to detect over 90% of CHD lesions in theory, in practice the fetal CHD detection rate is closer to 50%. Prior literature suggests a key cause of this startling diagnosis gap is suboptimal acquisition and interpretation of fetal heart images. DL is a novel data science technique that is proving excellent at pattern recognition in images. DL models are a function of the design and tuning of a neural network architecture, and the curation and processing of the image data used to train the network. Preliminary Studies — We have assembled a multidisciplinary team of experts in echocardiography and CHD (Drs. Grady, Levine, and Arnaout), DL and data science (Drs. Keiser, Butte and Arnaout), and statistics and clinical research (Drs. Arnaout and Grady) and secured access to tens of thousands of multicenter (UCSF and six other centers), multimodal fetal imaging studies. We have created a scalable image processing pipeline to transform clinical studies into image data ready for computing. We have designed and trained DL models to find key cardiac views in fetal ultrasound, calculate standard and advanced fetal cardiac biometrics from those views, and distinguish between normal hearts and certain CHD lesions. Hypothesis — While DL is powerful, much work is still needed to adapt it for clinical imaging and to translate it toward clinically relevant performance in patient populations. We hypothesize that an integrated ensemble DL/ML approach can lead to vast improvements in fetal CHD diagnosis. Aims — To this end, the main Aims of this proposal are (1) to develop and optimize neural network architectures and efficient data inputs to relieve key performance bottlenecks for DL in fetal CHD; and (2) to deploy DL models population-wide to evaluate their ability to improve diagnosis, biometric characterization, and precision phenotyping over the current standard of care. Our methods include DL/ML algorithms and retrospective imaging analysis. Environment and Impact — This work will be supported in an outstanding environment for research at the crossroads of data science, cardiovascular and fetal imaging, and translational informatics. The work proposed will provide valuable tools and insight into designing and evaluating both the data and the algorithms for DL on imaging for clinically relevant goals, and will lay important groundwork for DL-assisted phenotyping for both clinical use and precision medicine research.