Generation of echocardiogram images for 3D image enhancement and localisation

超声心动图(Echo)是评估和处理所有心脏疾病的关键。超声心动图是用超声波来制作心脏的运动图像。在临床使用的功率水平下，声波的使用是无痛和无害的，产生声波所需的设备成本低且便携，这提供了方便[波特]。超声心动图的这些优点有助于它今天在临床上的广泛应用。三维超声心动图(3D Echo)可以量化绝对心腔体积，可视化心脏的3D结构和动态运动图像，特别是心脏瓣膜结构[Shiota]。它为传统的二维超声心动图提供了更多的临床信息，并已被确定为连续量化接受化疗的癌症患者左心室容量和射血分数的最佳超声心动图方法。此外，与2D版本不同，3D ECHO不依赖于平面定位，不需要几何建模，也不对心腔的形状做出假设[程]，进一步提高了其重复性和准确性。然而，与今天的2D成像技术相比，3D回波的主要局限性一直是，而且仍然是图像质量较差。目前，空间分辨率受到探测器可以发送和接收的波束和扫描数量的限制。这明确地限制了横向分辨率，结果也削弱了图像对比度。更高分辨率的3D图像将使计算突出腔体积时更准确，并改进该方法的可视化能力。改善图像质量的尝试主要集中在对传感器本身的更改[CAAS]。超声心动图的一个重要问题是，如果没有足够的训练和经验，它很难使用，特别是在经食道超声心动图(TEE)成像中。可能很难知道探测器的位置，也很难理解到底拍到了什么。探头的自动定位将通过帮助用户识别探头在身体内的姿势来简化执行TEE的过程。此外，自动本地化为执行回声本身创造了更多自动化的空间。例如，机器人执行器可以在回声的特定阶段被调用，以执行与医疗专业人员相比更适合的任务。为了解决这些问题，我们计划开发一条管道，根据2D回波图像输入确定探头在体内的位置和方向。这将通过首先开发一种算法来实现，该算法在给出2D回波图像时，识别其在心脏3D解剖模型中最可能的位置。一旦开发出来，我们将使用这个算法来预测探头在体内的位置和方向，从而开发出使用2D切片支持TEE成像的3D图像制导的方法。最后，在DPhil的最后部分，我们将使用这条管道来支持利用更高质量的2D回波图像来增强3D回波图像，从而克服3D TEE的主要障碍之一。在整个项目中使用的方法将主要基于最先进的机器学习工具，特别是卷积神经网络。该项目是与GE Healthcare合作进行的，属于EPSRC医学成像研究领域。参考文献：Potter，A.，Pearce，K.和Hilmy，N.(2019)。超声心动图在初级保健中的益处。《英国全科医学杂志》，第69期(684)，第358-359页。Https://doi.org/10.3399/BJGP19X704513 Shiota，T.(2008年)。三维超声心动图：现状和未来。心脏病杂志，52(3)，169-185。Https://doi.org/10.1016/J.JJCC.2008.09.004

Echocardiography (Echo) is key to the assessment and management of all cardiac diseases. Echocardiograms are produced using ultrasound waves to create a moving picture of the heart. At the power levels used in the clinic, the use of sound waves is painless and harmless, and the devices required to generate them are low cost and portable which provides convenience [Potter]. These advantages of echocardiography contribute to its widespread clinical use today. 3D Echocardiography (3D Echo) allows the quantification of absolute cardiac chamber volumes and visualisation of the 3D structure and dynamic motion images of the heart, especially heart valve structures [Shiota]. It offers significant additional clinical information to traditional 2D echo, and has been identified as the best echocardiographic method for sequential quantification of left ventricle volumes and ejection fractions in patients with cancer undergoing chemotherapy. Moreover, unlike the 2D version, 3D echo is not reliant on plane positioning, does not require geometric modelling and does not make assumptions about the shapes of the chambers of the heart [Cheng], furthering its reproducibility and accuracy. However, the main limitation of 3D echo has been and still is the inferior image quality compared with today's 2D imaging technology [Lang]. Currently, the spatial resolution is limited by the number of beams and sweeps that the probe can send and receive. This limits lateral resolution specifically and as a result, also weakens image contrast. Higher resolution 3D images would enable more accuracy when calculating salient chamber volumes and improve the visualisation capabilities of this method. Attempts to improve image quality have mostly focused on changes to the transducer itself [Casas]. An important issue with the echo modality is that it is difficult to use without significant training and experience, especially in transoesophageal echocardiography (TEE) imaging. It can be difficult to know the probe's position and understand what exactly is being imaged. Automatic localisation of the probe would ease the process of performing TEE by helping the user identify the probe pose within the body. Furthermore, automatic localisation creates scope for more automation in performing the echo itself. For example, robotic actuators could be called on at certain stages of the echo to perform tasks that they would be better suited to compared to a medical professional. In order to tackle these issues, we plan to develop a pipeline that determines the probe's position and orientation, within the body, from 2D echo image inputs. This will be achieved by firstly developing an algorithm that, when given a 2D echo image, identifies its most likely location in a 3D anatomical model of the heart. Once developed, we will use this algorithm to predict the probe's position and orientation within the body, and hence develop methods to support 3D image guidance of TEE imaging, using 2D slices. Finally, in the final part of the DPhil we will use this pipeline to support the leveraging of the higher quality 2D echo images to enhance the 3D echo images, overcoming one of the main obstacles to 3D TEE. The methodologies used throughout the project will be mostly based on state-of-the-art machine learning tools, in particular Convolutional Neural Networks. This project is undertaken in partnership with GE Healthcare and falls within the EPSRC Medical Imaging research area. References Potter, A., Pearce, K., & Hilmy, N. (2019). The benefits of echocardiography in primary care. British Journal of General Practice, 69(684), 358-359. https://doi.org/10.3399/BJGP19X704513 Shiota, T. (2008). 3D echocardiography: The present and the future. Journal of Cardiology, 52(3), 169-185. https://doi.org/10.1016/J.JJCC.2008.09.004