Enhanced x-ray angiography analysis and interpretation using deep learning

增强的X射线血管造影术分析和 利用深度学习进行口译 美国每年进行100多万次诊断X射线血管造影术以指导冠心病的治疗 动脉疾病(CAD)，花费超过120亿美元。尽管这是临床护理的标准，但视觉解读 容易出现观察者之间和观察者内部的变异性。最近，作为NHLBI支持的未来多中心的一部分 评估胸痛的影像研究(PROMISE)试验，我们的研究小组表明，心脏病专家 超过19%的血管造影术为阻塞性CAD(血管狭窄程度大于50%)。考虑到中心地位 对于制定治疗计划的血管造影解释，准确性的降低可能会导致不必要的 糟糕的结果和增加了我们医疗体系的成本。考虑到潜在的影响是巨大的， 仅在美国，口译准确率提高1%，每年就能使1万多名患者受益。 因此，我们团队正在开发一种深度学习驱动的X射线血管造影分析系统(DeepAngio) 提高医生解释能力的技术。在第一阶段，超过1,000张血管成像的Promise数据集是 用于构建基于卷积神经网络(CNN)的深度学习模型。我们达到了0.89的面积 在与专家一起识别图像中的阻塞性CAD的接收工作特征(AUROC)下 真实得分(超过我们建议的第一阶段里程碑&gt；0.85AUROC)。 现在，在第二阶段，我们提出了一种创新的图像学习管道，将解剖学和时空学结合起来 来自视频序列的信息(类似于心脏病专家阅读器)。完整的端到端X射线血管造影术视频 将开发处理流水线，并在10,000名血管造影正常和 分级异常CAD。我们的基于补丁的框架分析模型将发展到基于CNN的全框架 血管造影图(左心与右心)的分类和冠状动脉的分割(LAD，Lcx， 和RCA)。由递归神经网络(RNN)实现的多帧分析方法将配备我们的 具有动态时间信息的模型，以准确地估计病变的存在。我们第二阶段的目标是 提高阅读的特异性，并将我们的第一阶段概念验证研究结果转化为临床 有意义的工具。由一组介入性心脏病专家进行的多读者、多病例评估 没有DeepAngio的预测将评估临床可用性，以改善冠状动脉狭窄的估计。 从长远来看，我们希望心脏病专家与DeepAngio作为辅助工具的结合将改善 血管造影解释的临床准确性。

Enhanced x-ray angiography analysis and interpretation using deep learning Over 1 Million diagnostic X-ray angiograms are performed annually in the US to guide treatment of coronary artery disease (CAD) and cost over $12 billion. Despite being the clinical standard of care, visual interpretation is prone to inter- and intra-observer variability. Recently as part of the NHLBI supported Prospective Multicenter Imaging Study for Evaluation of Chest Pain (PROMISE) trial, our research team showed that cardiologists misinterpreted over 19% of angiograms obstructive CAD (greater than 50% vessel stenosis). Given the centrality of angiographic interpretation to the development of a treatment plan, reduced accuracy can lead to unnecessary poor outcomes and increased costs to our healthcare system. The potential impact is significant given that increasing interpretation accuracy by 1% could positively benefit over 10,000 patients each year in the US alone. Thus, our team is developing an X-ray angiographic analysis system (DeepAngio) driven by deep learning technology to enhance physician interpretation. In Phase I, the PROMISE dataset of over 1,000 angiograms was used to build our Convolutional Neural Network (CNN) based deep learning model. We achieved a 0.89 Area Under the Receiving Operating Characteristic (AUROC) for identifying obstructive CAD in images with expert scored ground truth (exceeding our proposed Phase I milestone of >0.85 AUROC). Now in Phase II, we present an innovative image learning pipeline to incorporate anatomical and spatiotemporal information from video sequences (similar to a cardiologist reader). A full end to end X-ray angiography video processing pipeline will be developed and tested in a new cohort of 10,000 patient angiograms with normal and graded abnormal CAD. Our patch-based frame analysis model will advance to CNN full frame-based classification of angiographic views (left heart vs. right heart) and segmentation of coronary vessels (LAD, LCx, and RCA). A multiple frame analysis approach enabled by a Recursive Neural Network (RNN) will equip our model with dynamic temporal information to estimate lesion presence accurately. Our goal for Phase II is to improve reading specificity and translate our Phase I proof of concept research findings into a clinically meaningful tool. A multi-reader, multi-case evaluation by a group of interventional cardiologists interpreting with and without DeepAngio predictions will assess clinical usability to improve coronary stenosis estimation. In the long term, we hope the combination of a cardiologist with DeepAngio as an assistive tool will improve the clinical accuracy of angiographic interpretation.