Synergistic Inverse Problems Omni-Solver for Expeditious High Quality Multimodal Cardiovascular MRI via Deep Compressive Sensing and Data Coalescing

心血管疾病是全世界最常见的疾病和死亡原因。在英国，它约占所有死亡人数的25%，每年给NHS造成的损失约为90亿英镑。医学成像被用来筛查和诊断疾病，并计划和监测治疗。心血管磁共振(CMR)是一种安全的技术，可以在不使用X射线的情况下对心脏的结构和功能进行详细的非侵入性成像，这可能会略微增加患癌症的风险。图像的获取方式--使用不同的“序列”(或多模式CMR)--非常灵活，可以改变图像的信息内容，以突出疾病的“生物标志物”。这些图像生物标记物可以更早地诊断疾病，更好地制定治疗计划。典型的CMR研究持续大约一个小时，可能需要几个小时以上的分析，通常需要报告的临床医生使用计算机鼠标手动识别和勾勒感兴趣的结构。虽然许多类型的图像可以在短时间内屏气采集，但覆盖范围更大的高细节图像可能需要5-10分钟才能完成，并且呼吸运动控制不佳会降低图像质量，影响图像生物标志物的可靠性和实用性。在保持或改善图像质量的同时提高采集速度，以及对生成的图像进行快速、可重复和全自动的分析，将导致新的生物标记的开发，并提高现有生物标记的可靠性。在本奖学金提案中，我将致力于通过使用先进的基于深度学习的信号处理方法，将多模式CMR成像速度提高12倍(取决于成像序列)。深度学习是一种新技术，它教计算机做人类很自然的事情：通过例子学习。这是通过使用快速的计算机与“大数据”并行工作实现的，并在不同的应用中产生了令人印象深刻的结果。然后，我将用一种新的“转移学习”技术来提高获得的CMR图像的质量，这种技术教会计算机通过使用它用来完成类似先前任务的模型来完成给定的任务。将设计一种“数据合并技术”，以利用现有的“大数据”。此外，我将开发和实现一个全自动的软件流水线来分析生成的图像，在~2分钟内提取心脏解剖和图像生物标志物(加速系数~100)。这些方法将使用现有的CMR图像进行开发，并将在我们的商用CMR扫描仪上实施，用于在健康志愿者和患者中进行前瞻性测试，这些患者包括心肌梗死(心脏病发作)、房颤(快速不规则心律)和先天性心脏病(出生缺陷)。使用新的“深度学习”方法，将CMR成像速度提高到12倍。发展创新的“转移学习”和“数据合并”技术，通过提高图像分辨率、抑制噪声和纠正模糊和其他伪影来提高图像质量。研究和实施全自动检测和分析软件包/流水线，用于评估新的和可靠的图像生物标志物。ii.Researchi的潜在好处。更快的成像和更好的图像质量相结合，将允许自动分析从CMR图像中提取的图像生物标记物。这将使处理时间减少约100倍，降低临床医生的可变性，并允许对心血管疾病患者进行更好的指导和评估。更快的成像将缩短研究持续时间并改善患者体验。它还将增加患者吞吐量，减少等待名单，并降低每项研究的成本。开发的技术将广泛适用于所有类型的CMR图像，也适用于其他患者群体，也适用于其他类型的医学成像。

Cardiovascular disease is the most common cause of illness and death worldwide. In the UK, it accounts for about 25% of all deaths, and it costs the NHS roughly 9 billion pounds each year. Medical imaging is used to screen and diagnose for disease and to plan and monitor treatment. Cardiovascular magnetic resonance (CMR) is a safe technique that allows detailed non-invasive imaging of the structure and function of the heart without using X-rays, which may increase the risk of cancer slightly. How the images are acquired - using different 'sequences' (or multimodal CMR) - is very flexible and can change the information content of the images to highlight 'biomarkers' of disease. These image biomarkers can allow earlier diagnosis of disease and better treatment planning.A typical CMR study lasts about an hour and can take several more hours to analyse, often requiring the reporting clinician to manually identify and outline structures of interest using a computer mouse. While many types of image can be acquired in a short period of breath-holding, highly detailed images with greater coverage can take 5-10 minutes to complete and image quality can be reduced by poor respiratory motion control which affects the reliability and usefulness of the image biomarkers. Increasing the speed of acquisition while maintaining or improving image quality together with a rapid, reproducible and fully automatic analysis of the resulting images will lead to the development of new biomarkers and improve the reliability of existing ones.In this fellowship proposal, I will work on methods to speed up multimodal CMR imaging by factors up to 12 (depending on the imaging sequence) by using an advanced 'deep learning' based signal processing approach. Deep learning is a new technique that teaches computers to do what comes naturally to humans: to learn by example. This is achieved by using fast computers working in parallel with 'big data' and has produced impressive results in different applications. I will then improve the quality of the CMR images obtained with a new 'transfer learning' technique that teaches a computer to accomplish a given task by using the model it used to do a similar previous one. A 'data coalescing technique' will be devised to take advantage of available 'big data'. In addition, I will develop and implement a fully automatic software pipeline to analyse the resulting images, extracting the cardiac anatomy and image biomarkers in ~2 minutes (speed-up factor of ~100). The methods will be developed using existing CMR images and will be implemented on our commercial CMR scanners for prospective testing in healthy volunteers and patients including those with myocardial infarction (heart attack), atrial fibrillation (fast irregular heart rhythm) and congenital heart disease (birth defect).I. Goals of Researchi. To speed up CMR imaging by acceleration factors up to 12 using new 'deep learning' methods.ii. To develop innovative 'transfer learning' and 'data coalescing' techniques to boost image quality by increasing the image resolution, suppressing the noise and correcting the blurry and other artefacts.iii. To investigate and implement a fully automated detection and analysis software package/pipeline for the assessment of new and reliable image biomarkers.II. Potential Benefit of Researchi. The combination of faster imaging and better image quality will allow automated analysis of image biomarkers extracted from CMR images. This will reduce the processing time by a factor of ~100, reduce clinician variability and allow better guidance and assessment of patients with cardiovascular disease.ii. Faster imaging will reduce study duration and improve the patient experience. It will also increase patient throughput, reduce waiting lists and reduce cost per study.iii. The techniques developed will be widely applicable to all types of CMR image, to other patient groups and also to other types of medical imaging.