Deep Learning Enabled Endovascular Stroke Therapy Screening in Community Hospitals

项目总结/摘要 中风是美国第五大死亡原因。血管内卒中治疗（EST） 彻底改变了大血管闭塞（LVO）急性缺血性卒中（AIS）的管理， 中风导致的不成比例的残疾虽然这种疗法已被证明可以显着改善 在多项临床试验中，这些研究几乎都需要筛查晚期乳腺癌患者。 神经成像，如CT灌注（CTP），这是大多数社区医院无法提供的一种方式。作为 因此，迫切需要一种工具，能够利用已经存在的基础设施， 在社区医院里。我们设想一种基于软件的服务， 使用CT血管造影（CTA）评估EST。我们开发并测试了一种新型深度神经网络的原型， DeepSymNet网络架构。我们的初步数据表明，独特地使用CTA，我们可以 确定（1）是否存在大血管闭塞（2）缺血核心的程度和（3） “处于风险”的组织体积（半影）高于或低于临床试验中使用的阈值，当 与使用CTP同时获得的结果相比。 我们将通过三个目标来实现我们的项目目标： - 目的1 -建立一个最大的多机构数据集，用于急性缺血性脑卒中的神经影像学研究。 中风我们将获得一个多中心数据集，包括来自全国15家医院的成像和临床数据。 德克萨斯州和加州，从一系列扫描仪、成像采集协议和医院类型（即大型 学术界和较小的社区）。 - 目标2 -开发可解释的深度学习模型，以确定EST的资格。我们将有条不紊地 测试一组模型架构、数据增强策略、损失函数和预处理步骤 在DeepSymNet上。我们将针对梗死核心和半影的各种定义来训练和测试该算法 基于CTP结果的体积。这种方法将允许模型适应不断变化的定义， EST资格。 - 目标3 -评估基于DeepSymNet的模型在大型多中心独立 数据集。为了实现这一目标，我们将在患者成像和数据上部署DeepSymNet软件， 多个医院，这些医院在软件的创建中没有使用。我们还将验证我们的方法， 通过验证CTA源梗死核心的盲读，单独使用CTA确定缺血核心 由专家阅读器对同时获得的CTP结果进行图像。 这些目标的完成将通过支持创建和 在绝大多数社区，决策支持工具的有效性可随时转移到患者床边 全国各地的医院。通过这样做，我们希望扩大高质量EST筛选的范围， 成千上万的AIS患者。

Project Summary/Abstract Stroke is the 5th leading cause of death in the United States. Endovascular stroke therapy (EST) has revolutionized the management of large vessel occlusion (LVO) acute ischemic stroke (AIS), which accounts for a disproportionate amount of disability in stroke. While this therapy has been shown to significantly improve clinical outcomes in multiple clinical trials, these studies nearly all required screening patients with advanced NeuroImaging such as CT Perfusion (CTP), a modality not available to the majority of community hospitals. As such, there is a pressing need to for a tool able to identify EST candidates leveraging the infrastructure already existing in community hospitals. We envision a software-based service to automate the NeuroImaging evaluation for EST using CT angiography (CTA). We developed and tested a prototype of a novel deep neural network architecture called DeepSymNet. Our preliminary data indicate that uniquely using CTAs we can determine (1) the presence or absence of a large vessel occlusion (2) if the extent of ischemic core and (3) volume of tissue “at risk” (penumbra) is above or below the thresholds used in the clinical trials, when compared to concurrently obtained results using CTP. We will pursue our project goal with three aims: - Aim 1 - Establish one of the largest multi-institution dataset for neuro-imaging research in acute ischemic stroke. We will acquire a multi-center dataset including imaging and clinical data from 15 hospitals across Texas and California, from a range of scanners, imaging acquisition protocols, and hospital types (i.e. large academic and smaller community). - Aim 2 - Develop interpretable deep learning models to determine the eligibility for EST. We will methodically test a set of model architectures, data augmentation strategies, loss functions and pre-processing steps based on DeepSymNet. We will train and test the algorithm against various definitions of infarct core and penumbral volume based on CTP results. This approach will allow for models adaptable to the everchanging definition of EST eligibility. – Aim 3 - Evaluate the external validity of DeepSymNet-based models on a large multi-center independent dataset. To accomplish this aim, we will deploy our DeepSymNet software on patient imaging and data from multiple hospitals, which were not used in the creation of the software. We will also validate our approach of using CTA alone to determine ischemic core by validating blinded reads of infarct core from CTA source images performed by expert readers against concurrently acquired CTP results. Completion of these aims will have a sustained, transformative impact by supporting the creation and validation of decision support tools readily translatable to the patient bedside in the vast majority of community hospitals across the country. In doing so, we hope to expand the access to high-quality EST screening to thousands of additional AIS patients.