AI-Enabled Assessment of Cardiac Function from Echocardiography

博士项目的目的：开发用于自动解释超声心动图图像的深度学习技术，并使用大规模数据集对其进行训练/评估。从超声心动图中自动表征心脏功能。自动评估瓣膜心脏状况和功能。基于机器学习的模型输入和输出质量控制项目描述：超声心动图是评估和诊断心血管疾病（CVD）的第一个呼叫中心。然而，对于许多临床生物标志物的准确和稳健的定量，需要心脏磁共振（CMR）成像。传统上，从CMR估计这样的生物标志物需要大量的专家交互，例如用于在心动周期上勾画左心室心肌的边界。我们最近发表的一项研究表明，使用最新的深度学习技术可以避免这种相互作用[1]。尽管如此，超声心动图仍然比CMR更便宜，更广泛。在这个项目中，我们的目标是翻译，调整和扩展我们为CMR开发的技术到超声心动图领域。这是一个合适的时间这样的项目-深度学习技术通常需要大量的数据进行训练和验证，最近已经有一些大规模的超声心动图数据库可用[2，3]。此外，我们还可以从KCL/GSTT Biobank获得数千份超声心动图扫描。我们相信，我们在开发基于自动机器学习的生物标记物评估管道方面的专业知识以及此类数据集的可用性，为高影响力的转化研究创造了令人兴奋的机会。之前使用深度学习进行超声心动图分析的工作主要集中在左心室（LV）内膜边界的自动分割[4-6]，以估计射血分数（EF）。然而，由于图像质量低、伪影的存在以及与不同病理相关的不寻常图像特征，边界识别容易出错。因此，这些算法可能缺乏鲁棒性。为了克服这一限制，一些工作集中在直接估计EF而不进行内边界分割[7-8]。虽然这种解决方案可能更可靠，但它的可解释性较低，临床医生更难以评估结果的准确性。我们建议开发基于机器和深度学习的技术，用于自动量化超声心动图扫描。我们将从基于CMR的模型中研究迁移学习的使用，并设计领域自适应技术，以利用我们在这项新任务中建立的知识。此外，我们将寻求超越简单的指标，如舒张末期和收缩末期容积和EF的估计，描绘更丰富的心脏健康和疾病的图片。此外，超声心动图仍然是评估心脏瓣膜疾病和反流的第一线技术，因为它具有出色的可视化瓣叶，这是在CMR中不可见的。该项目的一部分将侧重于对不同心脏瓣膜的解剖结构和功能进行自动评估。通过适当的质量控制工具和置信度测量，这些技术原则上可以在毫秒内工作，并在扫描时向超声医师提供实时反馈。随后对大规模估计的生物标志物进行分析，可以对CVD的性质和心脏疾病的进展进行有趣和有价值的研究。

Aim of the PhD Project:Develop deep learning techniques for automated interpretation of echocardiography images and train/evaluate them using large-scale datasets.Automated characterisation of cardiac function from echocardiography.Automated assessment of valvular heart condition and function.Machine learning based quality control of model inputs and outputsProject Description:Echocardiography is the first port-of-call for assessment and diagnosis of cardiovascular disease (CVD). However, for accurate and robust quantification of many clinical biomarkers cardiac magnetic resonance (CMR) imaging is required. Traditionally, estimating such biomarkers from CMR has required a significant amount of expert interaction, e.g. for contouring the boundaries of the left ventricular myocardium over the cardiac cycle. We have recently published a study that demonstrated that much of this interaction can be avoided using the latest deep learning techniques [1].Nevertheless, it remains the case that echocardiography is cheaper and more widely available than CMR. In this project we aim to translate, adapt and extend the techniques we have developed for CMR into the realm of echocardiography. This is the right time for such a project - deep learning techniques normally require a large amount of data for training and validation, and recently a number of large-scale databases for echocardiography have become available [2,3]. Furthermore, we have access to thousands of echocardiography scans from the KCL/GSTT Biobank. We believe that our expertise in developing automated machine learning based pipelines for biomarker estimation, and the availability of such datasets, creates an exciting opportunity for high-impact translational research.Previous work on using deep learning for echocardiography analysis has mainly focused on automated segmentation of the left ventricular (LV) endocardial boundary [4-6] for estimation of ejection fraction (EF). However, boundary identification is prone to errors due to low image quality, the presence of artefacts, and unusual image features linked to different pathologies. As a result, these algorithms can lack robustness. To overcome this limitation, some works have focused on the direct estimation of EF without endocardial border segmentation [7-8]. Although this solution could be more reliable, it is less interpretable and more difficult for clinicians to assess the accuracy of the results.We propose to develop machine and deep learning-based techniques for automated quantification of echocardiography scans. We will investigate the use of transfer learning from our CMR-based models, and design domain adaption techniques to take advantage of our established knowledge in this new task. Furthermore, we will seek to go beyond the estimation of simple metrics such as end-diastolic and end-systolic volumes and EF, to paint a much richer picture of the heart in health and disease.In addition, echocardiography remains the first line technique for assessing valvular heart disease and regurgitation due to its excellent visualisation of the valve leaflets, which are not visible in CMR. Part of this project will focus on the automated assessment of the anatomy and function of the different heart valves. With appropriate quality control tools and confidence measures, the techniques could, in principle, work in milliseconds and give the sonographer real-time feedback whilst scanning. The subsequent analysis of the estimated biomarkers at scale could enable interesting and valuable research into the nature of CVD and the progression of the heart into disease.